Machine Learning For Beginners

Complete AI for Beginners with Real-world

Business Applications
Declan Mellor
© Copyright 2020 by – Declan Mellor - All rights reserved.
This document is geared towards providing exact and reliable information in regards to the topic and
issue covered. The publication is sold with the idea that the publisher is not required to render
accounting, officially permitted, or otherwise, qualified services. If advice is necessary, legal or
professional, a practiced individual in the profession should be ordered.
The information provided herein is stated to be truthful and consistent, in that any liability, in terms
of inattention or otherwise, by any usage or abuse of any policies, processes, or directions contained
within is the solitary and utter responsibility of the recipient reader. Under no circumstances will any
legal responsibility or blame be held against the publisher for any reparation, damages, or monetary
loss due to the information herein, either directly or indirectly.
Table of Contents
MACHINE LEARNING FOR BEGINNERS
INTRODUCTION
INTRODUCTION
THE PURPOSE OF THIS BOOK
WHAT IS ARTIFICIAL INTELLIGENCE?
HOW IS MACHINE LEARNING USED?
RECENT IMPROVEMENTS IN DATA ANALYSIS
IMAGE RECOGNITION
VOICE RECOGNITION
MEDICINE AND MEDICAL DIAGNOSIS
STOCK FORECASTS
LEARNING ASSOCIATIONS
FINANCE
SPAM DETECTION
INTRODUCTION TO STATISTICS
CHOOSING THE RIGHT TYPE OF MODEL FOR MACHINE
LEARNING
LEARNING UNDER SUPERVISION
REGRESSION ANALYSIS
DECISION TREES
RANDOM FORESTS
CLASSIFICATIONS
LOGISTIC REGRESSION / CLASSIFICATION
K NEAREST NEIGHBORS
VECTOR SUPPORT
KERNEL SUPPORT VECTOR
NAIVE BAYES
LEARNING WITHOUT SUPERVISION
CLUSTERING
K-MEANS CLUSTERING
DIMENSIONALITY REDUCTION
NEURAL NETWORKS
LEARNING REINFORCEMENT
Q LEARNING
SEMI-GUIDED LEARNING
ENSEMBLE MODELLING
THINGS YOU NEED TO KNOW FOR MACHINE LEARNING
DATA
PREPARE THE DATA
PROGRAMMING TOOLS
DEVELOP MODELS
ANACONDA FOR PYTHON
ALGORITHMS
VISUALIZATION TOOLS
MORE ADVANCED THINGS THAT ARE USEFUL
EPILOGUE
Introduction
Congratulations on purchasing Machine Learning for beginners and thank
you for doing so.
There are many possibilities in the field of machine learning. It is used as a
tool by almost every major industry. Whether you are interested in
healthcare, business and finance, agriculture, clean energy and many others,
there is someone who uses the power of machine learning to make their job
easier.
Unfortunately for these industries, but luckily for you, there is a huge
shortage of talent in the field of data science and artificial intelligence.
While entry-level data science jobs remain competitive, there is a
significant shortage of experienced data professionals who can perform the
high-level functions. It is a newer field in computer science, with a younger
group of individuals making up for much of the field.
It can be very financially rewarding if you manage to get a job in data
science. In 2016, the average data scientist made about $ 111,000, with
predicted growth over the next five years. About half of the data scientists
working in the field have a Ph.D. It is not a requirement, but it is something
to think about if you are looking for a real career as a data scientist.
If you want to add machine learning to your wheelhouse so that you can
better understand and implement it in your own company or projects, a
Ph.D. may not be necessary. But for those looking to enter the field, higher
education is recommended as it will help you stand out from the field.
Indeed.com called machine learning the best career in 2019 and it's easy to
see why. With a huge demand for talented data scientists and a lucrative
payoff, it's worth checking out. And big data doesn't seem to be
disappearing anytime soon with increasing connectivity and internet use by
consumers and businesses more than ever. Data is part of our modern
world, and as the complexity and size of data increases, even more
specialized knowledge and skills will be required to complete the task.
To supplement the knowledge in this book, I strongly recommend seeking
more knowledge in statistics and programming. A good foundation of
statistical knowledge is required to perform any work in machine learning,
because statistical mathematics provides the structure and justification for
all models and algorithms that data scientists use for machine learning.
The purpose of this book
This book is not intended to be a comprehensive machine learning
textbook. Instead, it gives you a foundation of knowledge to continue your
studies of machine learning and artificial intelligence. To continue your
studies and master the subject, a great deal of study has to be done.
Discusses the general structure and organization of machine learning
models, the general terms and basic statistical concepts necessary to use and
understand machine learning.
To be a data scientist requires an understanding of statistics and quantitative
analysis. After all, artificial intelligence and machine learning are rooted in
statistics. This provides the anchor and foundation for the type of math
required.
While coding is not required to understand this book, it is an important part
of machine learning. In order to process large amounts of data, data
scientists must have working knowledge of computer programming to "tell"
the data what they want them to do. This book doesn't provide much
information about coding, but it will provide resources and ways to get you
started on coding your own. In any case, by the end of the book I will help
you set up Python with the necessary libraries and toolkits to help you learn
how to program.
The most used language in machine learning is Python. It is a versatile
language that is relatively easy to learn and freely available. Python
packages are designed for data analysis to speed up your encoding. C ++ is
also fairly common, but more difficult to master. A third option is R, which
is quite popular because it is free and open source. Students often use it for
its availability and simplicity. The downside to using R is that it can't
handle huge data sets commonly used in machine learning and artificial
intelligence, which is somewhat limiting.
Machine learning computers are distinguished for not only remembering
new information, but applying it to new situations in the future. There is a
difference between remembering and learning. There is an important
distinction between giving a code line to a computer and creating a machine
learning model.
The basic feature of machine learning is the use of artificial inductive
reasoning. Artificial inductive reasoning means that a specific event gives
you reason to generalize a characteristic. This apple is green; therefore, all
apples must be green. But here you can see why inductive reasoning in
itself is not always perfect and why it is difficult to train computers to
follow the same thought process. A given piece of data is not necessarily
representative of thousands of other possible pieces of data. Therefore,
when we use
statistics and machine learning, we need to use enough data to reason with
confidence, without drawing the wrong conclusion from data that is
misinterpreted and misleading.
There are things we do every day as human beings that we consider
"common sense." These kinds of intuitive decisions cannot be explicitly
programmed into a computer, because the variables that help us make our
decisions are too difficult to measure. We probably don't need to see a
thousand different combinations of chess pieces on a chessboard to think
and plan ahead when we get into a situation we haven't seen before. We as
humans need much less data to be able to derive and learn.
This is where machine learning comes into play. In these situations where
the variables are complex and directions cannot be explicitly stated, they
need to be learned. Back to the example of a computer that can play
checkers. It would take far too long to teach someone to play checkers by
giving them all possible moves and all possible countermoves. Instead, you
teach someone the basics, and through play, the person learns what helps
them win and what doesn't.
Likewise, it is impossible to tell your computer all possible situations in a
drafts game and then tell the computer what to do in each situation. There
are far too many options. Instead, you should provide the computer with
enough data so that it can respond accordingly even when faced with a new
situation.
Another example we'll talk about later in this book uses artificial neural
networks to sort whether a photo is an image of a cat or a dog. As a human,
this type of classification would be too easy. We know what a dog looks
like because we've seen dogs, and we know what a cat looks like because
we've seen cats.
But there is no way to explicitly tell a computer how to tell the difference
between a cat and a dog. Instead, give the computer a set of training data
with pictures of cats and dogs, and tell the computer which pictures are cats
and which pictures are dogs. Ultimately, the model should be able to tell
you if new, unseen photos include a feline or a canine tooth.
The problem with explicitly programmed instructions is their inability to
change. If I tell a computer exactly what to do, using a programming
language with explicit instructions, the program will do that job very well,
but only that job. It doesn't change when it gets new information and it
doesn't change the method if it doesn't work properly
Over time, a machine learning model will be able to change itself as the
data changes so that it continues to adapt and remain accurate in a changing
unaccompanied environment. This offers a huge advantage because it
makes our models more adaptable to changes that are constantly around us.
Without machine learning and artificial intelligence, our computers would
have no way to keep up.
What is artificial intelligence?

In the 1950s, individual researchers started developing the first machines

for artificial intelligence. Previously, only small experiments had been
conducted with artificial intelligence; especially when cracking codes
during World War II. It was an emerging field and only a few people
seemed to be aware of the potential at first
Now artificial intelligence is used in a variety of applications across
multiple sectors, including problem solving, learning, planning, reasoning
and logic. It enables computers to perform tasks that normally require
human thinking. To 'think' like a person, computers need data from which
they can learn.
Artificial intelligence has an almost mythical place in people's minds. I bet
if you said the words artificial intelligence to most people, there would be
images of robots walking around like humans. This kind of science
fictionalization of artificial intelligence makes people wary when they hear
the term. But it is not as scary as it sounds. It has done a lot of good in
medicine and business, transportation and communication. While
impressive progress has been made in the field, the delusion of a conscious
computer is still a long way off. Yet the emergence of self-driving cars and
computers and telephones that can talk stirs the imagination.
While it sounds more like something from science fiction, artificial
intelligence is now in so much of our daily lives. If it sounded spooky to
you at first, let me remind you of all the technology that artificial
intelligence has brought to our lives in recent years.
The last time you turned on Netflix, you browsed through a list of the
shows you watched and the movies you watched again. It turned that list
into data and created another list of recommendations. It predicted movies
you would enjoy based on what you already liked. This is done through
machine learning, a subset of artificial intelligence.
If you have a smartphone, you can use voice commands to search for things
hands-free. You tell your phone that you are looking for cafes in your area
and your phone says "Search cafes in your area." Within a few seconds, a
list of results will appear and you don't even have to type anything. It
recognized your voice and understood what you were saying. This is part of
natural language processing, another subset of machine learning. Every
time you open your email account and you tag spam; your email host will
learn how to better identify spam. This is another type of machine learning.
Artificial intelligence is therefore not necessarily conscious robots that want
to take over as we know it. From now on, it is much more favorable than
that. It is also extremely useful and it can teach things for which we cannot
explicitly program it.
Artificial intelligence requires something called artificial reasoning, also
known as machine reasoning. When people learn new things and draw
conclusions, we go through a process known as inductive reasoning. We
take pieces of information to draw new conclusions. Usually there is no
hard and fast rule to teach. We learn from experience and draw our own
rules through cumulative experience. For example, I could tell you
that it snowed 15 times last December. Therefore, it will snow again this
December. Every day in January was cold, so every day in January will be
cold. So I have to bring a jacket.
We weren't born wired to know that snow would fall in December, or that
January would be cold. We learned these things through experience and
used inductive reasoning to generalize about future December and January
periods. Based on our inductive reasoning, we make the logical decision to
prepare and bring a jacket next winter.
The experiences we had with snow in December and cold in January
represent our "data". These are the input of our environment from which we
constantly learn.
People think differently than machines because we don't interpret numerical
data patterns. We learn from positive and negative rewards and from the
feelings we experience in our daily lives. By having a computer use
inductive reasoning, we get closer to "human-like" machines
So to learn computers, they need data to learn from. Data usually needs to
be numeric so that it can be interpreted by mathematical models and
algorithms. If we give a computer enough data, it will create the parameters
to design its own model or algorithm, to predict new situations based on
previous experience. This is the basis of machine learning. Power the
computing experience so that it can predict new results in the future through
inductive reasoning.
Artificial intelligence is particularly interesting because computers are
already better than humans for some tasks. They can draw mathematical
conclusions about a dataset with thousands of inputs in seconds. No human
on earth could process that kind of information so quickly. If we could use
machine learning to examine data from a complex data set of 100 variables,
we could probably learn more about trends and patterns that were very
complex and difficult to distinguish manually. This is what makes
computers such a useful tool and why they have contributed to massive
advances in data science. Using computers for data analysis makes it easier
to find patterns and matches that you don't even know exist or that may not
even have been considered.
Computers perform very poorly in other tasks. Some of these tasks seem
very simple to us. Like identifying the difference between a photo of a cat
and a photo of a dog. But for a computer, this is extremely complicated to
figure out. Therein lies the current challenge with artificial intelligence.
Bridging that gap between the type of inductive reasoning that people can
do and the type of reasoning computers is good at that.
Key concept: inductive reasoning. Using information from experiences and
our environment to draw generalizations.
The ability to tell the difference between these photos based on our
knowledge of cats and dogs; this is what we know as reasoning. The goal of
artificial intelligence is to teach computers how to have the same skills as
human reasoning
Computer models have been used to process natural languages. Natural
language processing gives computers the ability to understand "natural"
languages, or what we know as our human languages. Natural language
processing is based on machine learning techniques to understand speech
and text and to respond to commands and interactions.
This technology is becoming very common and accessible. GPUs (graphics
processing units) are becoming more widely available and cheaper, which
means that data sets are getting bigger and the use of machine learning is
increasing. You may have used it when talking to Siri on your iPhone.
When you say something to Siri, your phone will receive the audio. To
interpret it, it turns your audio into text. Your phone then analyzes the text
to derive meaning from the command you gave it.
Natural language processing is one of the most common uses of machine
learning and we use it every day. When we use a web search function, we
use natural language processing. Translation apps need to take our voice or
our text and analyze the sentence structure to gain meaning. When you type
a paper or word document, your word processor uses natural language
processing to look for grammatical and spelling mistakes.
Despite its popularity, it is a very complex field of computer science and
artificial intelligence. Being able to interpret the meaning of the alphabet in
an almost infinite amount of combinations requires vast amounts of data for
the computer to understand what you are writing or saying.
Besides being able to understand what we say and write; computers can
also make strategic decisions based on what they learned from data in the
1990s. IBM created a computer called Deep Blue that defeated a world
chess master. It was the first computer to perform such a task. Because of
the simplicity of the rules in chess, computer scientists at IBM chose to
train their computers to play. But there are thousands of possible moves and
arrangements that the pieces can take on once the game has started. The
computer had to learn this using data.
What makes machine learning unique for other forms of computer science
is the ability of models to change their methods over time to adapt new
data. What distinguishes a machine learning model from a regular line with
explicit code is that machine learning absorbs new data and improves itself.
It can also perform tasks that require planning and contain strategic
components. The Deep Blue computer had to be adept at analyzing possible
strings of motion, rather than just one motion at a time.
The same technologies that allowed a computer to beat a world chess
champion now allow self-driving vehicles to get a passenger safely from
point A to point B. Compared to the relative simplicity of chess, self-
driving cars must plan and interpret hundreds of variables to keep the
passenger safe. It goes beyond the two-dimensional data analysis used by
chess machines. Self-driving cars must master multi-dimensional data
analysis to navigate the ever-changing environment on the road.
The machine learns through trial and error, repeats the task over and over,
and learns from failures and successes. These experiences are introduced as
data and over time the machine will know the probability of failure or
success with every possible movement.
Machine learning models interpret potential conditions in the environment.
For an algorithm that plays chess, these are all possible moves and all
possible moves by the competitor. The algorithm is an amalgamation of
goals and possible actions. By using this data, a plan is drawn up to
optimize the chances of achieving the goals. It also allows computers to
self-learn without specific prompts through programming.
Trying to get a computer to do all these things sounds simpler in theory than
in practice. Most of the features we just mentioned; from dams to self-
driving cars require advanced statistical techniques to optimize the outcome
and train a machine that knows how to 'win' with a high degree of accuracy.
Machine learning falls under the larger umbrella of artificial intelligence.
Artificial intelligence is a branch of computer science that includes
reasoning, natural language processing, planning and machine learning. The
term was first coined by a computer scientist named John McCarthy in
1956. You will also hear the term data science, which includes artificial
intelligence and machine learning. Data science is a broader term, but is
often used to describe machine learning. Machine learning experts are often
referred to as data scientists, both in this book and beyond. There is an
overlap between data science and machine learning, but it is not the same.
Data science is more of a general term, while machine learning is part of
data science.
How is machine learning used?
Machine learning is a popular buzzword today. You have often thrown
around the term machine learning, especially in data science for digital
marketing. Other familiar terms such as artificial intelligence, data science
and data mining may seem synonymous with machine learning. There are
minor differences between these different fields, all of which fall under the
umbrella of data science.
Data science is the management and analysis of data, and within data
science there are many ways to analyze the data and use it to learn. Machine
learning is his own field within computer science. The idea is to predict
something and, as you add more data, compare predictions to actual
outputs. Over time, your forecasting ability should improve and errors will
decrease.
One of the most important functions of the human brain is the ability to
change our behavior based on the results of past events and situations. If a
situation has a positive outcome, we remember that, and in turn, if a
situation has a negative outcome, we also store it in our memory. Later we
will use this "data" to make decisions about new ones
situations. Over time, we learn to interpret situations, even if they are
completely new, and we are not quite sure how to behave or react.
Machine learning helps us create a mathematical way to copy human
learning ability over time and through a new experience. Machine learning
models learn over time and improve their prediction methods, improving
the outcome. Past data is collected and over time the model can make better
and more accurate predictions. Over time, the program will be able to make
more accurate predictions because of the new data it has received. It learns
over time how it can be done better at completing a certain task.
Part of it has inspired science fiction and fueled fears that artificial
intelligence will surpass us and take over the world. Ultimately, our
machines will be able to do everything we can, only better. Our computers
will surpass and leave us. Despite general fears, artificial intelligence is still
a relatively young field with a long way to go. While it may not take over
the world anytime soon, machine learning has changed the job market today
and will continue to change it on its own in the future.
Jobs that used to require human thinking can now be done with machine
learning. Factories, medical diagnosis and even taxis can be run with
artificial intelligence and machine learning. Data is becoming an
increasingly important field. Patterns in data cannot be interpreted with a
human brain, so we use machines to detect patterns.
Machine learning is used in the financial sector to detect fraud. Algorithms
can now detect when a financial transaction has fraud characteristics.
Businesses can spot fake reviews by recognizing word patterns and timing
for previously fake posts.
Our phones use speech recognition to understand what we are saying and
respond to our requests. Social media use complex data analysis to
recognize patterns in our photos and to see who is in a photo before we start
tagging. All this is done by data collection and interpretation through
machine learning. The patterns in photos that tell who appears are data that
is analyzed and improved by machine learning models.
In 1959 Arthur Samuel first popularized the term 'machine learning'. As a
graduate of MIT, he chose to create a computer program that could play
checkers. He chose checkers because of the relative simplicity, but also
because checkers contain many possible strategies. This predated IBM's
deep blue, but the theory was the same. Over time, let the computer learn
with new data.
The computer looked at the positions of each chess piece and every possible
movement. Each move had a score with the chance of winning. The
algorithm included factors such as the number of pieces on the board and
the number of pieces that were cononed.
The algorithm has memorized any combination of positions it has
experienced in a process called distance learning. For each of these
combinations, it remembered the score of the chance of winning. In a few
years, Samuel played numerous games with the machine to teach him how
to play.
Machine learning was created to express what the machine did to learn how
to beat opponents. The machine had to learn to play checkers and learn how
to reduce its mistakes to win. It was the first time that machine learning was
discussed as its own independent field of study outside of computer science
or artificial intelligence.
Sandford University calls machine learning "the science of making
computers act without being explicitly programmed." Samuel's computer
did not need to be programmed to remember every possible move made by
a king. Instead, the "experience" the machine gained served as data,
learning from each game to optimize its winning strategy. Statistics provide
the structure and mechanics for machine learning. Despite the birth of the
term machine learning in 1959, it has only been recognized as a separate
field in computer science since the 1990s with the advent of Deep Blue.
Normal computer programming uses an input command to control the
model. An example of this is connecting 4 + 4 to your python window. It
gives you the answer, 8. Instead machine learning uses the so-called input
data. Input data is what the machine needs to learn, while input command
would mean the machine is not self-learning. The programmer does not
specify what he wants the answer to be. Instead, the machine interprets the
data itself, making it self-learning. In machine learning, the machine takes
data over time and uses it to create a new model for something.
The machine detects patterns and structure within data based entirely on
statistical logic. It is entirely based on mathematical algorithms. Instead of
using intuition to search for patterns, patterns are found on a quantitative
and logical level by the machine learning model. The more relevant the data
to which a machine is exposed, the better it understands and can predict the
outcome of the model.
Machine learning is only useful if you can improve the efficiency of the
task you're trying to accomplish. Before starting your data, think
specifically about what you are trying to learn. What is the question you are
trying to answer with your machine learning model?
The environment is always changing, and if you want to create models that
are up-to-date and work in the changing environment, you need to keep
learning your model with new data. If you want your data to be useful to
you, you need to find a way to keep your data up to date with current
questions.
Learn as much as possible about your market or the environment in which
you operate. What things are you trying to understand? If you use machine
learning to improve your business, find out what types of variables are your
customers' choices. These are the things you want to identify and study in
your models. You need to understand your topic well before you can use
machine learning to study it. It is impossible to predict the future, even with
machine learning at your disposal. But if you can adapt your models to the
ever-changing data, you have a better chance in the long run than if you
stagnate.
Before you begin, you must have sufficient knowledge of the data you
choose. Where does it come from and how was it collected? In what format
is it and what are the challenges in interpreting it? These are the types of
questions you should be asking when you start, and I hope this book lets
you take a critical look at the data before you start.
It is not a comprehensive approach to predicting the future, and it will not
give you the magic formula to anticipate future business trends or stock
prices. But it is an incredibly useful tool that, when used properly, can make
decision making much easier.
The possibilities to experiment with machine learning are growing. Finding
the right people to fill the jobs needed to develop machine learning and
artificial intelligence is a challenge today. It is a specialized field and there
is a shortage of people with statistical and IT knowledge to help move the
field forward. This means that there is an opportunity for people who have
the skills needed for machine learning.
Recent improvements in data analysis
Machine learning has changed dramatically since its beginnings in 1959,
when Arthur Samuel's checkers played computer. But it has changed more
in the past two decades than in its entire history, especially with the
improvement of computing power. In the past, machine learning and big
data analysis were very limited. Only larger companies with expensive
technology could use data to make business

decisions.
Now almost anyone can use a certain amount of data for business or other
purposes with a laptop or home computer. Data is much easier to find and
the machines for processing it have also become much more accessible.
What used to cost expensive computing power can now be done much
cheaper and faster.
The advent of cloud technology has made it easier for smaller companies to
access large data sets without the need for huge amounts of data storage.
Now machine learning has become a completely different field of computer
science, with people specializing in machine learning and data science as
their own field.
Nowadays more and more things are connected and the internet is getting
bigger. This means that access to data is increasing, but data sources are
also changing. Even people's cars have computers inside them, which
means they create data that can be interpreted while driving. The vast
majority of Americans own a mobile phone and shop the web and use apps
for navigation. People use their phones to control household appliances,
which is another potential data source. There are Fitbits and smartwatches
that allow people to track health data.
The more devices, not just computers and telephones, but all kinds of
devices connected, the greater the possibilities for collecting and studying
data. This connection of everything; smartphones, smart cars, etc. make
people nervous that they risk losing their private data. They fear that their
privacy is at stake and that someone will always look at them. But machine
learning and data analysis make our lives much easier. Finding the right
products is easier, navigating is easier, and finding new music is easier. This
is all thanks to machine learning.
Image recognition
One of the applications of machine learning models is for sorting and
classifying data. This type of classification can even be used for the
classification of images. Search engines use this kind of algorithms to
identify photos, and social media sites now use facial recognition to identify
a person in a photo before the photo is even tagged. They do this by
learning from data composed from other photos. If your social media
account can recognize your face in a new photo, it's because it created
models with data from all the other photos in your account.
Image recognition techniques require in-depth learning models. In-depth
learning models are created with an artificial neural network, which will be
discussed in more detail later in this book. Deep learning is the most
complex type of machine learning where data is filtered through several
hidden layers of nodes. They are called hidden layers because the models
are unattended, which means that the features identified by the model are
not pre-chosen by the data scientist. Usually the features are patterns that
the model identifies on its own. Functions identified in neural networks can
be quite complicated, the more complex the task, the more layers the model
will have. Image sorting models may have only two or three layers, while
self-driving cars have between one and two hundred hidden layers.
We have made great strides in this area in recent years due to the increased
availability of computing power. Imagine the computing power needed to
run thousands of data points through hundreds of stacked nodes at once.
Deep learning and artificial neural networks have become more feasible
over the past decade with the improvement of computers and the reduction
of costs for processing large amounts of data. Certainly with the advent of
the cloud, which gives data scientists access to enormous amounts of data
without using physical storage space.
There is a website called ImageNet, a great resource for data scientists
interested in photo classification and neural networks. ImageNet is a
database of images that is publicly accessible for use in machine learning.
The idea is that by making it publicly available, improving machine
learning techniques will be a collaboration with data scientists around the
world.
The ImageNet database has approximately 14 million photos in the
database, with over 21,000 possible class groups. This offers a world of
opportunities for data scientists to access and classify photos to learn and
experiment with neural networks.
Every year, ImageNet hosts a competition for data scientists worldwide to
create new image classification models. The competition gets tougher every
year. Now they are starting to move to classifying videos rather than
images, which means that the complexity and required processing power
will continue to grow exponentially. Using the millions of photos in the
database, the ImageNet competition has made groundbreaking advances in
image recognition in recent years.
Modern photo classification models require methods that can be classified
very specifically. Even if two images have to be placed in the same
category, they can look very different. How do you make a model that can
distinguish them?

Take these two different photos of trees, for example. Ideally, if you were to
create a neural network model that classified images of trees, you would
want your model to categorize both as photos of trees. A person can
recognize that these are both photos of trees, but the characteristics of the
photo would make it very difficult to classify them with a machine learning
model.
The less differences the variables have, the easier they can be classified. If
all your photos of trees looked like the image on the left, with the tree in
full view with all its features, the model would be easier to make.
Unfortunately, this would lead to overfitting and when the model is
presented with data with photos like the right, your model may not classify
it correctly. We want our model to be able to classify our data even if they
are not that easy to classify.
Incredibly, ImageNet has been able to create models that classify data with
many variables and very similar data. Recently, they have created image
recognition that can even identify and categorize photos with different dog
breeds. Imagine all the variables and similarities that the model would need
to recognize in order to properly see the difference between dog breeds.
The challenge of identifying similarities between a class is known as intra-
class variability. If we have an image of a tree stump and a photo of a tree
outlined in a field, we are dealing with variability within the class. This
problem is how variables within the
the same class can differ, making it more difficult for our model to predict
which category they will fall into. Most importantly, a lot of data is needed
over time to improve the model and make it accurate.
To have an accurate model despite high levels of variability within the
class, we will have to use additional techniques with our neural network
models to find patterns between images. One method involves the use of
convolutional neural networks. Instead of just having one model or
algorithm, data is fed through several models stacked on top of each other.
The neural networks convert image features into numerical values to sort
them.
Unfortunately, it would be beyond the scope of this book to try to
understand the way these deep neural networks work, but there are many
books available that deal with these types of models and also include more
extensive explanations of the coding required to perform this type of
analysis. .
Voice recognition
Improvements in artificial intelligence have made speech recognition more
recently very useful. Most of our smartphones now have a certain level of
speech recognition, which means machine learning. Speech recognition
takes the audio data we give it and turns it into text that can be interpreted.
The difficulty of speech recognition is the irregularities in the way people
speak. Such as variability within the class. You and I may have different
accents and different inflections that are difficult to explain when teaching a
computer how to understand the human voice. If we both say the same
word with different accents, how do we teach the model to understand us?
Speech recognition also uses neural networks to interpret data, such as
image recognition. This is because the patterns in audio data are unlikely to
be recognizable to a human. Data scientists use sampling to interpret data
and make accurate predictions, despite the differences in people's voices.
Sampling is done by measuring the height and length of the sound waves,
which believe it may or may not be used to decipher what the user is
saying. The recorded audio is converted into the wave map of frequencies.
Those frequencies are
measured by numerical values and then passed through the hidden layers of
the neural networks to search for patterns.
Medicine and medical diagnosis
Machine learning is not only useful for digital marketing or to get
computers to respond to your requests. It also has the potential to improve
the medical field, especially in the diagnosis of patients using data from
previous patients.
With as much potential as machine learning has for medical diagnosis, it
can be challenging to find patient data available for machine learning due to
patient privacy laws. It is gradually gaining acceptance in the field of
medicine, meaning data is becoming available to data scientists.
Unfortunately, until now it has been difficult to have enough meaningful
data to make models related to medical diagnosis. But the technology is
there and available to use.
Machine learning can use image recognition to diagnose X-rays by bringing
data from different patients to imaging scans to make predictions about new
patients. Clustering and classification can be used to categorize different
types of the same disease so that patients and medical professionals can
better understand the variation of the same disease between two patients
and their survival rate.
Medical diagnosis with machine learning can reduce doctors' diagnosis
errors or give doctors something to provide them with a second opinion. It
can also be used to predict the likelihood of a positive diagnosis based on
patient factors and disease characteristics. One day, medical professionals
may be able to view data from thousands of patients about a particular
disease to make a new diagnosis.
But medical diagnosis is just one of many ways machine learning can be
used in medicine. Medical datasets remain small today, and the science of
machine learning still has a lot of untapped potential in the field of
medicine.
Stock forecasts
Stock traders look at many variables to decide what to do with a stock
whether they want to buy or sell or wait. They look at certain features of a
stock and trends in the market environment to make an informed estimate
of what to do. It has been done this way for years. Brokers and traders had
to do manual research to make the best estimate.
Machine learning can now be used to do the same, except that machine
learning can be done much faster and more efficiently. To be an effective
trader you need to be able to analyze trends in real time so you don't miss
out on opportunities. Machine learning can help traders find agreements
between stocks to make financial decisions using statistical data.
Traders can use linear regression models to study data on trends in past
stock prices and what variables cause a stock price to go up and down. They
can use these regressions to decide what to do with a stock.
Traders who want to analyze the performance of stocks often do this by
using a so-called support vector machine. A supporting vector machine is a
classification model in which data points are separated by a boundary line,
with a category on one side and another category or another. Traders will
use support vector machines to classify which stocks to buy and which
stocks to sell. Using certain variables that should be indicative of the
performance of a particular stock, that stock is placed on the side of the
boundary line that indicates whether the price is likely to rise or fall.
Deep learning is also often used when making stock models. The hidden
layers of a neural network can be helpful in identifying unseen trends or
characteristics of a stock that could cause them to rise or fall in price.
There is no such thing as a certain bet or a risk-free investment. This was
true when people made decisions, and it is still true when we use data
science to make financial predictions. It is important to remember that
investing in the stock market will always be risky. It is impossible to create
a model that predicts anything reliable about the stock market. It is wild and
unpredictable. But we have already learned that machine learning can find
patterns that people may not be able to find on their own.
If you understand that stock market trends can be completely arbitrary and
unpredictable, it is helpful to have another model that allows you to
estimate the predictability of stocks. Knowing how accurate your
predictions are for a given stock is just as important as the predictions
themselves. Create a separate model to measure the predictability of a
particular stock so you know how reliable your predictions are. Different
stocks have different predictability levels. It is important to illustrate that
with your model so that you can choose from the most reliable predictions.
Traders will proceed with the final decision on whether or not a stock will
go up or down in value. But data science and machine learning can
streamline the information analysis process that will aid the decision-
making process. That's why you see more and more examples of machine
learning models used when predicting inventory, and why at least
familiarize yourself with the idea.
Learning associations
Marketers in all areas, from brick and mortar stores to online stores, are
always looking for ways to connect products and increase sales. Whether
you own a small bike shop or a huge online warehouse, finding patterns in
your customer's buying behavior will help you make proactive decisions to
boost sales and make more money.
Most of us will visit a supermarket during a certain week. Supermarkets are
a perfect example of using product positioning to generate sales. Each
supermarket will organize itself so that similar items are placed together.
Baked goods have their own aisle, while fruits and vegetables have their
place. They do this for two reasons; it makes it easier for the shopper to find
what he needs and improves the customer experience. Product positioning
can also help put customers in touch with products that they want to buy but
weren't looking for when they first walked into the store.
In addition to placing the vegetables in the same aisle, there is yet another
strategy that supermarkets can implement to lead customers to certain
products. They can derive characteristics from a customer buying a specific
product and use it to recommend other unrelated products. For example,
you can assume that someone who buys fresh vegetables from the vegetable
aisle will eat healthier. You can put vegetable smoothies in the same
refrigerator where you store fruit. If a customer is looking for craft beer,
you can tempt him with a snack and place the kettle chips on the same
island as 12 packs of light beer.
If all that makes sense to you, you're on your way to understanding a
technique called collaborative filtering. It is a machine learning technique
widely used in internet marketing. If your search data shows that you have
visited airline tickets to Cancun, you may see swimwear advertisements in
your browser.
Marketing data scientists always try to answer this question; how can we
use data to find a way to link a product to its target group? The point is to
use data to link two otherwise unrelated products together to drive sales.
It is a way of making recommendations to a customer based on what you
know about him. Machine learning can often find similarities or buying
patterns with customers we may not have been looking for. This is a
powerful marketing tool that is starting to emerge in modern times.
Previously, most marketing agencies had to use intuition to find their target
markets. Data scientists can now use quantitative data to paint a more
accurate picture of their ideal customer. If you are interested in using
machine learning in digital marketing, this is a subject you should know.
Collaborative filtering is different from just promoting a similar product to
a customer. You make predictions about a customer's taste or preferences
based on data you have collected from other customers. You base this
prediction on a correlation you found between two products, and then a
measure of the likelihood that product Y will be bought with product X.
You use these estimates to decide what to market and to whom.
Spotify uses a similar process when making song recommendations. It uses
data from all the music that you liked over time. If there is a connection
between two artists, which means that many people have both artists in their
library, the model can predict the probability that you will like the other
artist.
The more products you have in your store, the more intensive it will be to
find these correlations. In a perfect world, you will look for correlations
between every different combination of products you have in your store.
This method of finding the probability that you like one product based on
buying another product is called the Apiori algorithm. There are three
criteria that must be met to confirm that there is a link between the two
products and that you must somehow link them in your store. The first
criterion is support. This gives you a measure of the popularity of a specific
product. Of all your transactions, how often does this item appear in
people's shopping cart?
The second part is confidence in the correlation between the two products.
How likely are customers to buy a Y product when they buy an X product?
After all, what is the lift of product Y? In other words, how likely someone
is to buy Y with X based on the popularity of Y alone.
The model can also use data from things like purchases, social media
engagements, etc. to make a prediction on the type of product you like. This
sets it apart as machine learning rather than just data analysis because the
model was looking for similarities, but the programmer didn't ask for a
specific output. Perhaps there are certain features or characteristics of the
group that the programmer is not even aware of. Perhaps unsupervised
machine data tells that there is a high correlation between the two types of
customers. These correlations take place all around us with similarities
between groups of people. Sometimes a good computer model is needed to
recognize the patterns in the data. Machine learning can find similarities
that would be impossible to see without the help of computers and good
models.
Data scientists in marketing industries are already using metrics to improve
their stores online, and if you want to track online retailing, it's a good idea
to start by reading about how data can help you identify similarities and
trends across products, with machine learning as your tool.
Finance
The financial sector is seeing an increase in the use of machine learning.
The use of data science and machine learning models makes the decision-
making process faster and more efficient for financial institutions. The
possibilities and applications of machine learning can be misunderstood,
which means that it is often underutilized or misused in the financial sector.
Work that was once tedious and required hundreds of hours of human work
can now be done by a computer in minutes. A well-known example is the
use of machine learning for call center and customer service work. Many of
the tasks that once required a human operator can now be accomplished
over the phone with a robot designed with machine learning.
In addition to customer service, banks can now process and analyze
contracts and financial information from thousands of customers that would
otherwise be labor intensive - used to prepare credit reports and predict the
likelihood of a customer defaulting on a loan. Machine learning techniques
can view a history of a borrower's transactions before the bank decides
whether to lend money to that person.
Machine learning is also used in fraud prevention. It has made the financial
sector safer. Machine learning has improved the bank's ability to detect
patterns in transactions indicative of fraud. Rather than being assigned
people to track the transaction and look for signs of fraud, machine learning
models can learn from fraud data to find patterns by automatically
searching millions of customer transactions.
Spam detection
A well-known example of a relatively simple machine learning tool is spam
detection. If we use guided learning and define the variables that are
relevant, the model will have certain characteristics to look for in received
email messages. The model may search for certain keywords or phrases to
detect whether an email is spam or not. Words like "buy" or "save" can let
your inbox know when you receive spam email. The problem with this
method is that in many cases there is not always spam. There may be other
keywords or phrases that we would overlook.
This is where strengthening learning comes in handy. There are so many
features that can be an indication of spam email, and some of them we
might not even be able to explain. Reinforcement learning will make it
possible to find these patterns independently, without explicit guidance.
Instead, we tell the model when it has correctly identified spam. Sometimes
we find in our inbox an email message that the model has not classified as
spam, so we manually move it to our spam folder. Now the model knows
that this message is spam, and this piece of data is added to the model to
improve the forecast next time. So over time, the machine gets better

the more relevant

data it gets.
This type of machine learning is known as classification. Our output falls
into separate categories. In statistics, discrete variables are variables that
can be identified in only a finite number of categories. An example of a
discrete variable is the number of cars a car dealer sells in a week. It is
discreet because the car dealer cannot sell half a car. The variable must be
an integer.
Introduction to statistics
Statistics is the mathematical science of data. It is a practice of collecting,
observing, and analyzing data to derive meaning and explore quantifiable
relationships between different variables. Machine learning is a form of
inferential statistics, which means that by investigating the relationship
between variables, we must be able to make predictions for new variables.
Statistics are used in a wide variety of disciplines. It is used in biology to
study and investigate the life of animals and plants. It has broad
applications in business, from making stock market forecasts to analyzing
consumer behavior. Economists use statistics to explain quantifiable
patterns in world markets. In medicine, statistics can be used to improve the
way doctors and disease specialists view the spread and prevention of
diseases.
Statistics are at the heart of machine learning. If you're not willing to dive
into statistics, machine learning is not for you. Machine learning uses
statistical algorithms to help computers learn. Machine learning is all about
tracking data and how computers can use data to improve themselves.
There are two types of statistics relevant to this book. The first is a
descriptive analysis, which you can use at the beginning of your modeling
process to look for indicators in your data. But most of what we do in
machine learning falls into another category called predictive analysis.
Important term; Descriptive analysis. The descriptive analysis helps us to
investigate that we are now. Looking at our current situation in the context
of the past and seeing why things are the way they are. Why do some things
sell better than others? What trends do we see in products currently on the
market?
Important term; Predictive analysis. Predictive analysis helps us to see
and understand what will happen in the future based on the current
indicators. When we use machine learning for predictive analysis, it is
important that we stay current and continue to provide the model with new
data. Which trends should we look out for?
Machine learning is just another way of understanding the data around us
and helping us understand our present and predict the future. But it requires
past and present data so we can find trends and see where they can lead.
Within statistics, there are two overarching categories of data that we will
use, and all of our data will somehow fall into one category or the other.
The first category is quantitative data. Quantitative data is data that can be
measured with a numerical value. Some examples of quantitative data are
length, income or the square footage of a house. All these variables can be
measured by a certain number, making them quantitative.
The second category is qualitative data. Qualitative data is data where the
variables are assigned to categories or classes. Examples of qualitative data
include a person's gender, blood type, whether a property has a pool or not.
This data can be sorted by identity and is non-numeric. Therefore it is
qualitative.
Quantitative data can be discrete or continuous. If we have a data set with a
variable that records the number of patients a hospital had last year, it
would be considered discrete. Discrete variables have a finite amount of
values they can have, and they are always whole numbers. It is impossible
to have half a patient or a percentage of a patient. Therefore, this variable is
discrete.
Data can also be continuous. An example of continuous data is an income
variable. Income can take half values, and there are almost endless
possibilities for the value of income in data.
Some other important terms to remember are the mean, median, and mode.
You will often hear these three things referred to in this book when we talk
about regressions. These are all different degrees of central tendency. The
mean is our mean value for data. If we have a variable for a person's age,
we find the average of the age by adding up all ages and then dividing by
the number of respondents in a dataset.
The median is the value in the middle of the dataset. If you had taken all the
answers by age and found the answer right in the middle of a sorted list of
answers, this would be your median.
The mode is the most common answer. If we took a sample of the ages of
eleven people and found that the ages were 19, 19, 20, 21, 22, 22, 22, 23,
24, 24, 25, the mode would be 22 because it is the most occurs in this
example. The median would also be 22 because it is in the middle of this
sorted list of responses.
When creating a statistical model, there are many important terms related to
the accuracy of our models. The most important and most mentioned in this
book are bias and variance. These are different types of prediction errors
that can occur when we create statistical models. Ideally, we want to
minimize the prevalence of bias and variance in our models. They will
always be there, and as a data scientist, you need to find the right balance of
bias and variance in your models, whether by choosing different data or
using different types of models. There are many ways to reduce variance
and bias within a model, depending on what you are trying to do with the
data. By reducing this with the wrong approach, you run the risk that your
model will be mounted too much or too little. If your model is biased, it
means that the average difference between your predictions and actual
values is very high.
Variance is how we can spread our predicted data points. Usually, the
variance results from overfitting on the sample data we used to create the
model. It does not accurately predict the outcome of new variables.
Errors will always occur in your models. It is a fact of life that no matter
how well you predict something, there is always a random or non-random

variation in the universe that makes

your prediction slightly different from the actual outcome.
I made a visual example of four bullseye targets to illustrate the difference
between high bias and variance models. In this case, the center of the
bullseye represents the true value that our model is trying to predict. The
top left corner is the ideal model. Note that all of our predicted data points
hit the bull's eye. This model is quite accurate and puts our predicted data
points around the true value. This is due to a low variance; a lack of
"staggered" data points and a low bias; substantiation that distorts our
results.

In the top right target, the model suffers from high variance. You can see
that our data points are grouped around the bullseye. Unfortunately, the
average distance between the predicted values and the bullseye is high due
to the large variance.
In the lower left target, the model did not suffer much from high variance.
The average distance between the predicted data points is low, but they are
not clustered around the bullseye, but slightly off due to high bias. This is
likely due to insufficient training data, which means that the model will not
perform well when introduced into new data.
The bottom right model has both a high variance and a high bias. In this
worst case scenario, the model is very inaccurate because the average
distance between predicted data points and the actual value is high and the
predicted data points are skewed.
Variance can be caused by a significant degree of correlation between
variables. Using too many independent variables can also be a cause of the
high variance. Sometimes, if the variance is too great, we can combat that
by allowing a small amount of bias in the model. This is known as
regularization. We will discuss that a little later.
In statistics, the population is the group of people or dataset you are trying
to analyze. The sample is the subgroup of that population, whose data you
use to create your model. The parameters are characteristic of the
population variables you are trying to identify and make predictions in your
model.
Descriptive statistics are the use of data to study a population. Descriptive
statistics typically include the mean or average, mode, media, size,
correlation. Machine learning falls into the category of inferential statistics
because we use the data to find patterns and relationships, but also to make
predictions based on this information. Inferential statistics or descriptive
statistics use the characteristics of your population to make predictions.
This is where your regression models and classification models come in.
When we derive something, we make a logical conclusion about a
population and the knowledge we gain.
If you look at data, also pay attention to the distribution. This is how the
data is distributed in our chart. It shows the frequency of values of our data
and how they appear in combination with each other.
We use our variance to find the standard deviation. Standard deviation is the
average of the distances between the predicted data points and the real data
points on a regression or prediction model.
We also need to make sure that we are aware of models that suffer from
over- and underfitting. An overfitted model is good at predicting results
using the training data, but if you are introducing new data, it will be
difficult. It is like a model that remembers instead of learns. It can happen if
you don't use random data in your training sample.
Underfitting describes a model that is too simple and does not investigate
significant data patterns. It may predict well, but the variables and
parameters are not specific enough to give us meaningful insights if you
don't have enough training data, your model may not fit.
One of the most common mistakes when looking at data is confusing
correlation with causality. If I told you that every person who killed last
year laid eggs every week, I couldn't claim that people who buy eggs are
murderers. Maybe I look at my data and see an increase in the number of
people buying milk as well as an increase in teenage pregnancies. Could I
argue that there is a connection between people who drink a lot of milk and
teenage pregnancies? Or teens who got pregnant caused people to buy more
milk.
This is the difference between correlation and causation. Sometimes the
data shows trends that appear to be related. When two events are correlated,
it means that they seem to have a relationship because they move along the
graph on a similar trajectory and in a similar space over time. While
causality means that the relationship between the two events causes one
event that causes another.
A number of criteria must be met to suggest that two cases are causally
related. The first is covariation. The causal variable, and the event, should
have created the need to be covariant, meaning that a change in one leads to
a change in the other.
The second criterion to be met is that the causative event must occur before
the event that it should have caused. For an event to be considered causal, it
must first come.
Third, the data scientist should check for external factors. To make it clear
that one causes the other; you must be able to demonstrate that the other
variables of the event are not the real cause. If the causal variable still
creates the effect even when other variables are considered, you can argue
that there is a causal relationship.
Choosing the right type of model for
machine learning

Imagine yourself as a carpenter. What kind of tools do you think you have
loaded in your truck when you arrive at a workplace? You will likely have a
hammer and a drill, as well as a few different types of saws. You probably
have a few planes and a good set of drills. If you know how to do your job,
you will know when you will know the purpose of each of these tools and
when to use them all. Each of these tools has a specific purpose. The drill
cannot do the work of a hammer, nor would you attempt to cut anything
with a hammer.
A data scientist who wants to do machine learning has their own set of
tools, each with a different purpose and designed for a different function.
Depending on the type of data you use and what you want to know, you
have to choose different algorithmic models to do the job.
Statistical algorithms can serve different purposes. Some predict a value; as
a regression model that predicts your income base based on your years of
education and work experience. Some models predict the probability of an
event, such as a medical model that predicts the probability that a patient
will survive a year or two, etc. Other models sort things by placing them in
different categories or classes, such as sorting photo recognition software
photo of different types of dogs.
Depending on the result you are looking for, you will need your statistical
tool belt. You must familiarize yourself with the technical skills of statistics.
You also need to know which tool to use and when to use it. Here I have
made an extensive list of the different types of statistical models that are
common in machine learning. To be able to write the code to build these
models yourself, I recommend that you take some time to study in the
programming language you have chosen. But this list gives you an
introductory understanding of each type of model and when they are useful.
In order for machine learning to be effective, you need to choose the right
model and the model that works best and have relevant data for the model
and demand.
Today, especially when using the Internet and digital marketing, there are
certain questions that cannot be properly understood without the use of data
and machine learning that can analyze it. Machine learning and data science
allow you to track your customers and their buying habits so you can better
adapt to their needs as they change.
The better you interpret your data, the easier it is to identify trends and
patterns, so you can anticipate the next change.
Machine learning can be divided into three different categories, each with
different unique algorithms serving different purposes. For starters, we'll
talk about the differences between guided, uncontrolled, and reinforcing
learning.
Learning under supervision

Programmers use labeled data in supervised learning. Before we start using

the algorithms, the data we look at is already predetermined. We know the
inputs and outputs we are looking for. X and Y. We are trying to find a
relationship between X and Y that we have chosen.
After you find a relationship between X and Y, you get a model that predicts
a result based on the relationships your machine has observed in the data.
Guided learning is used for regression and classification models. In machine
learning, we refer to characteristics as a certain measurable property or
characteristic of the data.
The first type of guided learning we will talk about and the first type of
statistical model is called regression. Regression is a model in which data
input and output are continuous. There are different types of regression, but
the most basic form is linear regression. We use linear regression to find a
relationship between an input X and an output Y. Once we have estimated
this relationship, we can predict Y with X. Linear models can, and usually
have, more than one X. In regression ; output Y has a numerical value.
Regression analysis
Regression is the simplest type of machine learning; this is usually where
you start when you first learn how to use your data. You have a set of X
values and you want to study their relationship to Y, the output. Our
independent variables, the X's in our model, gain weight and for each value
of X it is multiplied by the weight until the concatenated function creates a
prediction for Y.
We can create a predictive model for Y using data we already know the X
and Y for. If we bring this information back, we get the weights of X. If we
have enough relevant data, we can eventually predict Y and be unfamiliar
with known values for X.
We draw our known Y and X values on a scatter plot, and our regression
model finds the "best fit" line through the data points. The regression line is
called a hyperplane. The slope of the line is called the slope.
We can measure the distance between the predicted value and the actual
data point, and we call this measurement deviation. Our goal in creating a
linear regression is to minimize the deviation in our predictions. The
smaller the difference, the deviation, the more accurate your model.
Most statistical models used in machine learning are rooted in this first
algorithm. Create a model that predicts a result by plotting our data points
along a line or in clusters. But the line isn't always straight and sometimes
the line doesn't show us the best fit.
An example of a non-linear regression function is the Sigmoid function.
The Sigmoid function creates an S-shaped curve. Rather than predict a
value, the Sigmoid function takes independent variables and produces a
probability between one and zero.
Simple linear regression. In simple linear regression, we study the
relationship between some predicted value Y and some predictor X. We call
our Y the dependent variable because it depends on the value of X. Our X is
known as the independent variable.
If you took algebra or pre-calculus in high school or college, you may
remember that the equation was a line;
Y = mX + b
If you were to draw this equation, you would have a graph that looks
something like this:
As you can see, the line shows that for each value of X there is a different
value of Y. You can predict the value of Y for each new value of X. In this
graph, the value of Y increases as the value of X increases. This is not
always the case.
This is the most simplistic regression, but it is important to understand how
it works, as we will continue to build on it from now on. Most of the
statistical analysis involves a plot as shown above, which predicts an output
dependent variable based on an input, the independent variable. This is an
example of guided learning because we specify the Y variable and the X
variable we use before we start modeling.
With almost all predictions, there will be more than one independent
variable that will determine our dependent variable. This brings us to our
next type of regression.
Multiple linear regression. In data science and most statistics tasks, this is
the most popular type of regression. With multiple linear regression we
have one output variable Y as before. However, the difference now is that
we will have multiple X's or independent variables that will predict our Y.
An example of using multiple linear regression to predict the price of real
estate apartments in New York City. Our Y or dependent variable is the
price of an apartment in New York City. The price is determined by X, our
independent variables such as square meters, distance to transport, number
of rooms. If we were to write this as an expression, it would look something
like this:
apt_price = β0 + β1 square feet + β2 dist_transport + β3 num_rooms
We take sample data, data we already have where we know our X's and
their Y's and we view them in a graph like this:

You can see that the values of X and Y don't create a perfect line, but there
is a trend. We can use that trend to make predictions about future values of
Y. So we create a multilinear regression and end with a line going through
the center of our data points. This is called the best fit line and it is how we
will predict our Y when we get new X values in the future.
The difference here is that instead of writing m for slope, we wrote β. This
comparison is much the same as if I had written
Y = b + m1X1 + m2x2 + m3x3
Except now we have labels and we know what our X's and our Y's are. If
you see a multi-line equation in the future, it will most likely be written in
this form. Our β is what we call a parameter. It is like a magic number that
tells us what effect the value of our X has on the Y. Each independent
variable has a unique parameter. We find the parameters by making a
regression model. Over time, with machine learning, our model will be
exposed to more and more data, improving our parameter and making our
model more accurate.
We can make the model by using training data with the actual price of New
York City apartments and the actual input variables of square meters,
distance to transportation and many rooms. Our model 'learns' to
approximate the price based on real data. Then, when we connect the
independent variable for an apartment with an unknown price, our model
can predict what the price will be.
This is guided learning using a linear regression model. It is checked
because we tell the model what answer we want it to give us; the price of
apartments in New York City. It learns how to more accurately predict the
price as it gets more data and we continue to evaluate its accuracy.
Ordinary least squares OLS will try to find a regression line that minimizes
the sum of squared errors
Polynomial regression. Our next type of regression is called a polynomial
regression. In the last two types of regression, our models created a straight
line. This is because the relationship between our X and Y is linear, which
means that the effect X has on Y does not change if the value of X changes.
In polynomial regressions, our model results in a line with a curve.
If we tried to use linear regression to fit a graph that has nonlinear features,
we would do badly to make the best fit line. Take the graph on the left, for
example; the scatter plot has an upward trend as before, but with a curve. In
this case, a straight line does not work. Instead, with a polynomial
regression we will make a line with a curve corresponding to the curve in
our data, like the graph on the right
The equation of a polynomial will look like the linear equation, with the
difference that there will be a polynomial expression on one or more of our
X values. For example:
Y = mX2 + b
The effect X has on Y changes exponentially as the value for X changes.
Vector regression support. This is another important tool for data scientists
and one that you should familiarize yourself with. It is most commonly used
in case of classification. The idea here is to find a line through a space that
separates data points into different classes. Vector regression support is
another type of guided learning. It is also used for regression analysis. It is a
type of binary classification technique that is not related to probability.
To support Vector Regression, all your training data falls into one category
or the other. You want to know in which category a new data point falls.
Your data is separated by a hyperplane into these two classes. When
creating a model for

the hyperplane, try to find a hyperplane that maximizes the distance

between the two classes. For example, in the following image, you have a
scatter plot where the data points can be divided into two different classes.
In this case, lines one and three can separate the data points into two
separate classes. For your model, however, you should choose rule two
because it maximizes the margin between the two classes so they are
clearer. The larger the margin, the better.
Ridge regression. This is a technique often used to analyze data suffering
from multicollinearity. Using cam regression correctly can reduce standard
errors and make your model more accurate depending on the characteristics
of your data.
Ridge regression can be useful when your data contains independent
variables with a high correlation. If you can predict an independent variable
by using another independent variable, your model runs the risk of
multicollinearity. For example, if you use variables that measure a person's
height and weight; these variables probably create multicollinearity in the
model.
Multicollinearity can affect the accuracy of your predictions. Consider the
type of predictive variables you use to avoid multicollinearity. The type of
data you use, as well as the collection method, can cause multicollinearity.
Chances are you have not selected a wide range of independent variables.
Your data points may look too similar because your choice of independent
variables is limited.
Multicollinearity can also be caused by too specific a model. You have
more variables than data points. If you have decided to use a linear model,
which has aggravated multicollinearity, you can try applying a cam
regression technique.
Ridge regression works by allowing a little bit of bias in the model to make
your predictions more accurate. This technique is also known as
regularization.
Another method is to improve the accuracy of the model by standardizing
independent variables. The simplest way is to change the coefficients of
some independent variables to zero to reduce complexity. However, we will
not simply set them to zero, but standardization implements a system that
rewards coefficients closer to zero. This reduces the coefficients, which
reduces the complexity of the model, but the independent variables remain
in the model. This will give the model more bias, but it is a tradeoff for
more accurate predictions.
LASSO regression. LASSO regression is another 'shrinkage' technique. A
very similar approach to ridge regression in that it encourages leaner,
simpler prediction models. In lasso regression, the model is a bit stricter
about lowering the value of coefficients. LASSO stands for the least
absolute crimp and selection operator.
Data on our scatter plot is reduced to a more compact point, such as the
average of the data. Like ridge regression, we use this when the model
suffers from multicollinearity.
ElasticNet regression. ElasticNet regression works by combining the
techniques of LASSO and ridge regression. The main goal is to try to
improve the LASSO regression. It is a combination of both methods for
rewarding lower coefficients in LASSO and Ridge regression. All three of
these models are accessible via the glmnet package in R and Python.
Bayesian regression. Bayesian regression models are useful when we have
insufficient data or data with poor distribution. These types of regressions
are made with probability distributions instead of data points, which means
that the graph will appear as a bell curve representing the variance with the
most common values in the center of the curve.
In Bayesian regression, the dependent variable Y is not a value but a
probability. Instead of trying to predict a value, we try to predict the
probability of an outcome. This is known as frequentistic statistics and
Bayes' theorem forms the basis for this type of statistics. Frequent statistics
suggest whether something will happen and the probability that it will.
When we talk about frequentistic statistics, we also include conditional
probability. Conditional probability involved events the outcomes of which
depend on each other. Every time you throw a coin it is an independent
event, which means that the previous coin toss does not change the chance
of the next coin toss. Throwing a toss is therefore not a conditional
opportunity.
Events can also be dependent, which means that the previous event can
change the probability of the next event. Say I had a bag of marbles and I
wanted to know how likely it is to get different colors out of the bag. If I
have a bag with 3 green marbles and 3 red marbles, and I draw a red
marble, the chance of drawing a red marble decreases on my next draw.
This would be an example of conditional probability.
Decision trees
One of the models that we will discuss later is called neural networks. They
are the most advanced forms of machine learning and are used for many
different purposes. I've associated this with decision trees because of how
often people turn to neural networks for classification problems when much
simpler models are available. Decision trees and the related random forest
models can be just as useful.
Despite the power of neural networks, they cannot be used for everything.
Fortunately, we have options, and the purpose of this book is to know what
your options are if you decide to build a model. The next place to look
when neural networks don't work are decision trees. Decision trees divide
data into subcategories with decision and leaf nodes in the form of a tree.
Decision trees have a number of advantages over neural networks
(discussed later in this chapter). For starters, neural networks need huge
amounts of data and powerful computers to process them. The advantage of
using a decision tree is that they are relatively straightforward, especially if
you compare them to neural networks. Unlike most models in this book,
they are very easy to read a decision tree, even for the untrained eye. This
makes them a good candidate when choosing a model to be presented to
stakeholders.
Decision trees are another form of guided learning, which means that we
label the categories we want to sort before creating the model. In some
cases, decision trees can complete regression tasks, but most often they are
used as classification models. When decision trees are used for regression,
the leaf nodes end in probabilities.
Decision trees start with a so-called root knot at the top of the tree. Then the
root node is split into two nodes after the root node. Nodes are individual
leaves in the tree, and the middle nodes are where decisions are made,
known as the decision nodes. The decision tree ends at the bottom in what
is called a terminal node, at the bottom of a branch, where the decision is
completed.
Ideally, the decision tree sorts the data quickly, layer by layer. That's why
we call the model "greedy" because the top nodes try to sort the data as
quickly as possible, so fewer layers are needed. Like neural networks,
decision trees often suffer from overfitting. A decision tree usually does not
work with other datasets because the sorting is so specific to each dataset.
Below is an example of a decision tree about whether or not to assign an
applicant to the applicant, determined by qualification factors. The basic
node is whether or not the applicant has a university

degree,
followed by the decision modes which all lead to a decision not to hire or
award an application. You can see from this decision tree why this type of
model is so specific to the specific data you are working with, because each
dataset has different qualifications and thus each dataset will be sorted
differently.
Random forests
Using only one decision tree on your model can limit the categories in which
the data is split and the outcome of the decisions. Because the decision trees
are "greedy", this means that certain categories are chosen to sort, which
means that other categories cannot be chosen either. But there is an easy way
around that. One way to diversify your decision trees and improve the
accuracy of your model is to use random forests.
If a real forest consists of several trees, that is exactly what a random forest
is. Instead of having just one decision tree, split the data into several
decision trees. If you only have one tree, models can often suffer from large
variance. Creating a random forest is a way to combat that in your model. It
is one of the best tools available for data mining. Any forest is as close as
possible to a prepackaged algorithm for data mining purposes.
In a random forest, all trees work together. The overall result of all trees is
usually correct, even if a few trees end with poor predictions. To make the
final forecast, the results of all trees are added together. Using votes from the
mean values of all trees gives us a definitive prediction.
Since we use comparable data, there is a risk of correlation between the trees
if they all try to do the same. If we use trees that are less correlated, the
model performs better.
Imagine if we bet on a coin flip. We have a hundred dollars each and there
are three choices. I can flip the coin once and the winner of that toss can
keep $ 100. Or I can flip the coin ten times and we bet ten dollars every
time. The third option is to flip the coin 100 times and bet a dollar on every
roll. The real expected result of any version of this game is the same. But if
you've done 100 tosses, you're less likely to lose all your money than if
you've only done one toss. Data scientists call this method bootstrapping. It
is the machine-student equivalent of diversifying a stock portfolio. We want
to have a model that gives us an accurate prediction. The more we split our
decision trees, the more accurate our data will be. But it is important that the
individual trees have little correlation with each other. The trees in the forest
must be diverse.
How do we prevent correlation in a random forest? Each tree first takes a
random sample from the dataset, so that each tree has a slightly different set
of data from each other. The tree chooses a trait that creates the most
separation between nodes, in a greedy process, just like individual trees.
However, in any forest, trees can only choose certain elements from the
general group of elements, so each tree is separated by different elements.
So the trees are not correlated because they use different functions to make
decisions about classification. In any forest, it is best to use at least 100 trees
to get an accurate view of the data, depending on the dataset you are
working with. In general, the more trees you have, the less your model will
become overfit. Random learning in a forest is called a "low-supervision
technique" because our outcome has been chosen and we can see the sorting
method, but it is up to each tree to categorize and separate variables by
attributes.
Classification models will tell us in which category something falls. The
categories are initially defined by the programmer. An example of a
classification model that any forest could use would be a model that
determines whether incoming emails should place spam in your "inbox" or
"spam" folder.
To create the model, we create two categories that our Y can fall into; spam
and not spam. We program the model to search for keywords or a specific
email address that may indicate spam. The presence of words such as "buy"
or "offer" will help the model determine whether the email falls into the
spam category or not. The algorithm records data and learns over time by
comparing its predictions to the actual value of the output. Over time, it
makes minor adjustments to its model, making the algorithm more efficient
over time.
Classifications
A few times in this book we have referred to classification models. Some of
the models we have already mentioned can be classified, but the following
are more supervised learning models used specifically for classification.
Classification requires labeled data and creates discontinuous predictions.
The graphs are non-linear for classification problems. There can be two
classes in a classification problem, or even more. Classification models are
probably the most widely used part of machine learning and data science.
The first type of classification is binary classification. In binary
classification, the data is classified into two categories, labeled 1 or 0. We
call it binary classification because there are only two possible categories,
and all of our data falls into one or the other.
But there are cases where we have more than two categories and for this we
use multi-class classification models. We also have linear decision
boundaries, separating data on either side of a line. Not all data can be
classified on both sides of a decision boundary.
The first image shows an example of a classification with a linear decision
boundary. In the second image, there are almost two classes, but they
cannot be separated linearly. In the third image, data points are mixed and
linear boundary classification is not possible. Depending on the type of data
you use, there are different model choices that are better suited for different
tasks
Logistic regression / classification
This method is used to classify dependent and categorical variables.
Logistic regression calculates probabilities based on independent variables.
It values the variables "Yes or No" to sort them. Usually used with binary
classification.
If you cannot separate the data into classes using a linear boundary, as in the
examples above, this is the method to use. It is one of the most common
types of machine learning algorithms. Not only does it sort into categories,
but it also tells us the probability that a category exists.

We denote this model by taking the odds function, where p is the

probability of an event;

And creating a formula called the logit

K Nearest neighbors
K closest neighbors are one of the simplest and most widely used data
classification methods. It is a form of supervised learning that is used for
both classification and regression, and it is also the most basic clustering
algorithm. Simply put, it's about taking a data point and placing it with the
most common and nearest group on the scatter plot.
In KNN a new data point is classified based on the average median value of
its neighbors K. The nearest neighbors of a new data point 'vote' for which
classification it falls. K is the number of nearest neighbors who vote in the
model. Set k to a number - this is the number of nearest data points that the
new data point will analyze to choose which point it fits. The proximity of
data points is measured using an Euclidean distance.
Take the following two images as an example. We have split our data into
two classifications; the white dots and the black dots. A new data point is
introduced, the triangle, and we would like to predict which classification it
belongs to.
In this model we chose K = 4. If you choose k = 4, the four closest data
points are analyzed. The most common class among the adjacent data
points is the class in which the new data point will be placed. In this case,
you can see in the image on the right that the four white dots are the closest
classification. Therefore, the new data point is classified in that class.

There are a few factors to consider when choosing the value for k. The
higher the number for k, the closer we get to the true classification of our
new data point. There is an optimal point where the value of k must stop
rising to avoid overfitting.
If you choose to use a number for K that is too low, chances are your model
will have a high degree of bias. If you use too high a number, the
computational power required to calculate the value is too expensive. You
may consider using an odd number when choosing a value for K, rather
than an even number. Using an odd number is less likely to encounter a
draw between classes that vote for a data point. Data scientists often choose
the number 5 as the default for k.
Using a large number for K will be very data intensive. Large datasets are
also difficult to use with KNN machine learning models. If you are using
larger data sets, you need to calculate the distance between hundreds or
perhaps thousands of data points. It also does not perform well when you
use this method on a model that exists in more than two dimensions. Again,
it has to do with the computing power required to calculate this distance
between many data points.
Vector support
Support vector is another type of classifier. It classifies using a hyperplane.
In general, we use a support vector model with smaller data sets, where it
performs reasonably well.
Kernel Support vector
While we'll touch kernel support vectors, they are later used to sort classes
that cannot be separated with a linear divider. The dividing line can take
many forms (linear, nonlinear, polynomial, radial, sigmoid).
Remind the classification of a linear boundary we just talked about, with the
classification looking something like the following image:

In this image, our data can be classified

by a straight line separating the two different data categories. It would be
helpful if data can always be separated in this way, but unfortunately, in
fact, it's not always that neat and tidy, usually you will have to separate the
data in a way that looks more like this:

In this example, the data cannot be separated by a linear boundary line. So

instead we have to use a technique called the kernel trick. It uses a measure
of agreement between data points to classify them.
Naive Bayes
Think of Bayes' theorem from the first section on guided learning. Naïve
Bayes models assume that predictors are independent. This model is easy to
use and useful in large data sets. It is often used to sort spam emails.
We use Baye's rule here. The idea of Bayes Rule is that by adding new,
relevant information to what we already know, we can update our
knowledge based on that new information. If we wanted to know the chance
of rain falling this afternoon, we can find out what percentage of the days it
rains per year. But then we found out it was raining this morning. How do
you think this affects the chance of rain this afternoon?
So our ability to predict the probability of something will improve as we
receive more information about the event.

Mathematically, Bayes' theorem is expressed as follows:

So we can classify new data points using Bayes' theorem. The way it works
is when we get a new data point, we calculate the probability that that data
point falls into a category based on the characteristics of that data point.
Learning without supervision

Unattended machine learning uses untagged data. Data scientists do not yet
know the output. The algorithm must discover patterns itself, where
patterns would otherwise be unknown. Find a structure in a place where the
structure is otherwise imperceptible. The algorithm independently finds
data segments. The model looks for patterns and structure in an otherwise
unlabeled and unrecognizable mass of data. Unsupervised learning allows
us to find patterns that would not be observable without computer scientists.
Sometimes huge data sets have patterns and it is impossible to search them
all to find trends.

This is good for researching consumer buying behavior so that you can
group customers into categories based on patterns in their behavior. The
model may discover that there are similarities in buying patterns between
different subsets of a market, but if you did not have your model to search
through these vast amounts of complicated data, you will never realize the
nature of these patterns. The beauty of unsupervised learning is the ability
to discover patterns or features in vast amounts of data that you wouldn't be
able to identify without the help of your model.
A good example of unsupervised learning is fraud detection. Fraud can be a
major problem for financial companies, and with large amounts of daily
users, it can be difficult for companies to identify fraud without the help of
machine learning tools. Models can learn to recognize fraud as tactics
change with technology. To tackle new, unknown fraud techniques, you
must use a model that can detect fraud under unique circumstances.
It is better to have more data when detecting fraud. Fraud detection services
should use a range of machine learning models to effectively fight fraud.
Use of both supervised and non-supervised models. It is estimated that
approximately $ 32 billion in fraudulent credit card activity will take place
next year, by 2020 fraud detection models will classify the output (credit
card transactions) as legitimate or fraudulent.
They can be classified based on a function such as the time of day or the
location of the purchase. If a trader usually sells around $ 20 and suddenly
has a sale for $ 8000 from a foreign location, the model will most likely
classify this trade as fraudulent.
The challenge of using machine learning for fraud detection is that most
transactions are not fraudulent. If there were even a significant number of
fraudulent transactions among non-fraudulent transactions, credit cards
would not be a viable sector. The percentage of fraudulent card transactions
is so small that it can create models that are skewed that way. The $ 8,000
purchase in a strange location is suspect, but is the result of a traveling
cardholder rather than fraudulent activity. Unattended learning makes it
easier to identify suspicious buying patterns such as strange shipping
locations and random jumps in user reviews.
Clustering
Clustering is a subgroup of unsupervised learning. Clustering is the task of
grouping similar things. When we use clustering, we can identify attributes
and sort our data based on those attributes. When we use machine learning
for marketing, clustering can help us identify agreements in groups of
prospect customers. Unsupervised learning can help us sort customers into
categories that we may not have created using machine learning. It can also
help you sort your data when working with a large number of variables.
K-means clustering
K-means clustering works the same way as K-nearest neighbors. You
choose a number before k to decide how many groups you want to see. You
continue to cluster and repeat until clusters are more clearly classified.
Your data is grouped around centroids, which are the points in your chart
that you chose where you want your data clustered. You choose them
randomly and you have k. Once you have introduced your data into the
model, data points are placed in categories indicated by the nearest center of
gravity, which is measured by the Euclidean distance. Then take the average
value of the data points around each center of gravity. Keep repeating this
process until your results stay the same and you have consistent clusters.
Each data point is assigned to only one cluster.
You repeat this process by finding the mean values for x and y within each
cluster. This will help you extrapolate the mean value of the data points in
each cluster. K-means clustering can help you identify previously unknown
or overlooked patterns in the data.
Choose the value for k that is optimal for the number of categories you
want to create. Ideally, you should have more than 3. The benefit associated
with adding more clusters decreases as the number of clusters increases.
The higher the value for k you choose, the smaller and more specific the
clusters are. You wouldn't want to use a value for k equal to the number of
data points, because each data point would end up in its own cluster.
You should know your dataset well and use your intuition to guess how
many clusters are suitable and what kind of differences there will be.
However, our intuition and knowledge of the data is less helpful if we have
more than just a few potential groups.
Dimensionality reduction
When you use dimension reduction, you shrink data to remove unwanted
functions. Simply put, you reduce the number of variables in a data set.
If we have many variables in our model, we run the risk of having
dimensionality problems. Dimensionality problems are problems that are
unique to models with large data sets and that can affect the accuracy of the
prediction. If we have many variables, we need larger populations and
sample populations to make our model. With so many variables, it is
difficult to have enough data to have many possible combinations to make a
well-fitting model.
If we use too many variables, we may also encounter overfitting.
Overfitting is the main problem that would make a data scientist think about
dimensional reduction.
We must choose data that we do not need or that are not relevant. If we
have a model that predicts someone's income, do we need a variable that
tells us what their favorite color is? Probably not. We can remove it from
our dataset. Usually it is not so easy to determine when to delete a variable.
There are some tools we can use to determine which variables are not so
important.
Principle Component Analysis is a method to reduce the dimension. We
take the old set of variables and somehow convert them into a newer set.
The new sets we created are called main components. There is a tradeoff
between reducing the number of variables while maintaining the accuracy
of your model.
We can also standardize the values of our variables. Make sure that they are
all valued on the same relative scale so that you don't blow up the
importance of a variable. For example, if we measured variables as a
probability between 0 and 1 versus variables measured with whole numbers
above 100.
Linear discrimination is another method of diminishing the dimension
where we combine attributes or variables, rather than removing them
altogether.
Kernel Principal Component is the third method of reducing dimensionality.
Here variables are placed in a new set. This model will be non-linear and it
will give us even better insight into the real parameters than original data.
Neural networks

Neural networks are a form of machine learning called deep learning. It is

probably the most advanced machine learning method, and to really
understand how it works may require a doctorate. You could write a whole
book about machine learning most technical type of model.
Neural networks are computer systems designed to mimic the path of
communication in the human brain. In your body you have billions of
neurons that are all linked together and travel up through your spine and
into your brain. They are connected by root-like nodes that send messages
one by one through the neurons all the way up to the chain until it reaches
your brain.
While there is no way to replicate this with a computer yet, we take the
principle idea and apply it to neural computer networks to replicate the
learning ability as a human brain teaches; recognize patterns and derive
information from the discovery of new information.
In the case of the neural networks, as with all our machine learning models.
Information is processed as numerical data via neural networks. By giving
numerical data values, we give it the ability to use algorithms to make
predictions.
As with the neurons in the brain, data starts at the top and works down,
separating it into nodes first. The neural network uses nodes to
communicate through each layer. A neural network consists of three parts;
Input, hidden and output layers.
In the image below, we have a visual representation of a neural network,
where the circles are each individual node in the network. On the left we
have the input layer; this is where our data comes in. After the data has
passed through the input layer, it is filtered through several hidden layers.
The hidden layers are where data is sorted by different characteristics and
functions. The hidden layers search for patterns within the dataset. The
hidden layers are where the 'magic' takes place because the data is sorted by
patterns that we probably wouldn't recognize if we sort it manually. Each
node has a weight that will help determine the meaning of the sorted
attribute.

The best use of these neural networks would be a task that would be easy
for a human, but extremely difficult for a computer. Think of the beginning
of the book when we talked about reasoning and inductive reasoning. Our
human brain is a powerful tool for inductive reasoning; it is our advantage
over advanced computers that can calculate a large amount of data in
seconds. We model neural networks after human thinking because we try to
teach a computer how to 'reason' like a human. This is quite a challenge. A
good example of a neural network is the example we mentioned, we apply
neural networks for tasks that would be very easy for a human, but very
challenging for a computer.
Neural networks can take a huge amount of computing power. The first
reason neural networks are a challenge to process is because of the amount
of data sets needed to create an accurate model. If you want the model to
learn how to sort photos, there are many subtle differences between photos
that the model must learn to complete the task effectively. This leads to the
next challenge, namely the number of variables required for a neural
network to function properly. The more data you use and the greater the
number of analyzed variables, meaning there is an increase in hidden
networks. At any given time, hundreds or even thousands of features are
analyzed and classified through the model. Take self-driving cars as an
example. Self-driving cars have more than 150 nodes for sorting. This
means that the amount of processing power a self-driving car needs to make
decisions in a fraction of a second, while analyzing thousands of inputs
simultaneously, is quite large.
When sorting photos, neural networks can be very helpful, and the methods
used by data scientists are improving quickly. If I showed you a picture of a
dog and a picture of a cat, you could easily tell me which was a cat and
which was a dog. But a computer requires advanced neural networks and a
large amount of data to learn the model.
A common problem with neural networks is overfitting. The model can
predict the values for the training data, but when exposed to unknown data
it is too specific for the old data and cannot make general predictions for
new data.
Suppose a math test is on the way and you want to study. You can
remember any formulas that you think will appear on the test and hope that
when the test day comes, you can simply connect the new information to
what you have already remembered. Or you can study deeper; learn how
each formula works so you can get good results even when conditions
change. An overfitted model is like memorizing the formulas for a test. It
does well if the new data is similar, but if there is a variation it doesn't know
how to adjust. You can usually tell if your model is overloaded if it
performs well with training data but poorly with test data.
When we check the performance of our model, we can measure it by its
cost value. The cost value is the difference between the predicted value and
the actual value of our model.
One of the challenges with neural networks is that there is no way to
determine the relationship between specific inputs and the output. The
hidden layers are called hidden layers for a reason; they are too difficult to
interpret or understand.
The most simplistic type of neural network is called a perceptron. It derives
its simplicity from the fact that it has only one layer through which data
passes. The input layer leads to one classifying hidden layer and the
resulting prediction is a binary classification. Remember that when we call
a classification technique as binary it means that it only sorts between two
different classes, represented by 0 and 1.
The perceptron was first developed by Frank Rosenblatt. It is a good idea to
familiarize yourself with the perceptron if you want to learn more about
neural networks. The perceptron uses the same process as other neural
network models, but you usually work with more layers and more possible
outputs. When data is received, the perceptron multiplies the input by the
weight it receives. Then the sum of all these values is connected to the
activation function. The activation function tells the input what category it
falls into, in other words it predicts the output.
If you looked at the perceptron in a graph, the line would look like this:
The line of the perception graph appears as a step, with two values, one on
each side of the 1. These two sides of the step are the different classes that
the model will predict based on the input. As you can see from the graph, it
is a bit rough because there is very little separation between the classes.
Even a small change in an input variable will cause the predicted output to
be a different class. It does not perform as well outside of the original
dataset you use for training because it is a step function.
An alternative to the perceptron is a model called a sigmoid neuron. The
main advantage of using the sigmoid neuron is that it is not binary. Unlike
perceptron, which can classify data into two categories, the sigmoid
function creates probability rather than classification. The image below
shows the curve of a sigmoid neuron

Note the shape of the curve around one, where the data is sorted with the
perceptron; the step makes it difficult to classify data with only marginal
differences. With the sigmoid neuron, the data is predicted by the
probability of falling into a particular class. As you can see, the line curves
are at one, which means that the probability of a data point falling into a
certain class increases after one, but it's just a probability.
Learning reinforcement

Reinforcement learning is our third type of machine learning. Like

unsupervised learning, no input is given. The reinforcement learning model
must discover for itself how it can be most effective. Next, the data scientist
suggests improving the model based on its ability to predict an outcome,
which is evaluated by looking at comparisons between forecast and actual
value. This is the most progressive type of machine learning and where
much of the machine learning will be done in the future.
Think of it as playing a game; over time you learn by winning and losing.
You learn to win by playing and the more you become familiar with the
game, the better you understand the mechanics of winning. Over time, the
data scientist provides the model with feedback on data collection and
processing. It receives a reward signal, so it knows when it predicts an
outcome correctly. The game "win" therefore gives positive feedback to the
algorithm. This is the type of machine learning used in games, robots,
navigation and self-driving cars.
Q Learning
In Q-learning, the model communicates with its environment to improve
itself. You start with a series of states. States are the things in the
environment that stand as obstacles and roads in your environment. Called
"S." In chess, this would be the way all of your pieces can move, as well as
where all of your opponent's pieces are located. These are states.
The possible moves are called "A." If you are a pawn, your possible moves
are one square forward. If you are a rook, your possible moves are in any
direction in a straight line. Q is the value of the model, which starts at 0. As
you play the game, Q goes up and down depending on the interactions with
the environment. With negative interactions, the score Q decreases. With
positive interactions, the score Q increases. The algorithm learns how to
move so it can optimize the number Q. It is random at first. Over time,
these random movements result in positive and negative effects on Q, and
the machine learns how each movement will affect the score of Q. It has to
play a lot of games to improve the way it plays over time. It is much easier
said than done to apply this process in real life.
SARSA State Action Reward State Action There is only a slight
difference between SARSA and Q. They work the same way to give the
model a reward response.
Deep Q network Deep Q is used when ordinary Q is not general enough.
When he sees new things he has never seen before, he doesn't know what to
do. Q learning cannot adapt itself for things he has never seen. Deep Q uses
a neural network.
Markov Decision Process has a set of possible states, a set of models, a set
of possible actions and a real value reward. This model learns through
interaction with the environment through continuous interaction.
DDPG Deeply deterministic policy development Another reward state
model functions as an actor and critic.
Semi-guided learning
This type of machine learning uses a combination of guided and
unaccompanied learning. Some data is labeled, while others are not. But
most of the time, most of the data is not labeled. Semi-guided learning can
be used for classification, regression and prediction.
Semi-supervised machine learning is useful because labeling everything is
too time consuming and can damage the ability to find new patterns.
Ensemble Modelling

We have learned that diversifying our trees can create a more accurate
prediction. But what if, instead of using different versions of the same
model, we just used different models? This is a common trick in machine
learning, also known as ensemble modeling. By combining information from
multiple different types of algorithms, we can improve the accuracy and
predictability of our model.
Ensemble modeling is all about the division and victory mentality. Different
models give us different insights about the data that may not be recognizable
for other models. By combining the information from different models, we
can learn even more about the truth of our data.
Ensemble modeling also helps minimize bias and variance in our
predictions. Individual models may contain prediction errors, but the sum of
all our predictions will be more accurate.
There are a few different methods of using ensemble modeling:
The first is to take the mode of your predictions. That is, take the value most
common with the models. Whichever prediction occurs most often or has the
highest number of "votes" is the prediction we choose.
We can also take the average of all predictions depending on the type of
models we have. The average of all predictions becomes our final prediction.
Our ensemble must also take into account the reliability of individual
models. The results of our models are given different weights, making some
predictions more important than others based on reliability.
How do we know what kind of models we want to combine? We already
know from this book that there are different types of models to choose from,
each with different capabilities and benefits.
A common pair of models use neural networks and decision trees together.
Neural networks provide us with new information and the decision tree
ensures that we have not missed anything.
In addition to the bootstrapping and bagging we discussed earlier, there are a
few other ways to do ensemble modeling. Data scientists use a so-called
bucket of models. Here they use different types of models to use with the
test data and then choose the one that performed best.
Another idea is called stacking. Stacking uses different types of models and
then uses all the results to give us a prediction that is a combination of all.
Data scientists like to use ensemble modeling because we can usually make
better predictions with a variety of models than with a single model alone.
The downside of ensemble modeling is that we lose some of our legibility.
Having multiple models at the same time makes interpretation more
difficult, especially if you want to share the data with stakeholders who have
no knowledge of data science.
We can also use different versions of the same model, such as how random
forests improve prediction with multiple versions of themselves - using
neural networks with different sets of nodes and different values for k, or
numbers of clusters to see how that changes the outcome of our prediction
and find out if there is an optimal value for k, or if there are any groups or
subgroups that we may have overlooked.
It doesn't do much if we already have a strong model. But if we combine a
few models with weaker forecast capabilities, it usually improves overall
accuracy.
Things you need to know for machine
learning

To be successful with machine learning, you need to have the right tools to
work, just as you should have skills and the necessary tools when building a
house. Below is a list of the materials needed to do machine learning.
Data
In order to work with your data, you must have enough data to divide it into
two categories; training data and test data.
Training data is the data that you initially use when building your model.
When you first create your model, you need to give it some data to learn
from. With training data you already know the independent variables and
their respective dependent variables. This means that for every input you
already know the output of your data. Based on this data, your model learns
to predict the output itself. Our training data gives us the parameters we
need to make predictions. This is the data our machine learns from.
Test data is the data the machine receives as soon as you are satisfied with
the model and see what it does in the wild. In this data, we only have the
independent variables, but no output. With test data, we can see how well
our model predicts a result with new data.
Your training data should contain most of your data; about 70%, while your
test data is the remaining 30%. To avoid bias, make sure that the data you
choose for training data and test data is completely random when you split
them. Do not choose which data you want to use; let it be random. Do not
use the same data for training and testing. Start by giving the training data
to the machine and investigate the relationships between X and Y, then try
to see how well your model performed.
The main question to consider during this process is whether your model
will still work when presented with new data. You can test this through
cross validation. This means that you will test your model for data that you
have not yet used. Have some data on hand that you have not used during
training to see how accurate your model is at the end.
You can also use K-fold validation to check the accuracy of your model.
This method is quite easy to use and generally unbiased. It's a good
technique to use if we don't have a lot of data to test with. For K-fold
validation, we will split our data into K-folds, usually between 5 and 10.
Test each fold and see how they performed over all folds when you are done
testing. Usually, the larger your number for k, the less biased your test will
be.
So far we've talked about models that interpret data to find meaning and
patterns. But what kind of data are we going to use? Where do we get our
data from and what does it look like?
Data is the most critical part of machine learning. After all, your model only
learns with data, so it is important that you have relevant and meaningful
data. It came in many shapes and sizes, different structure depending on the
types of data. The more structured the data, the easier it is to work with.
Some data have a very small structure and these data are more difficult to
interpret. Face recognition data can be huge and has little meaning to the
untrained eye.
Structured data is better organized. This is the type of data you are likely to
use when you first start. It will help get your feet wet and you can start
understanding the statistic involved in machine learning. Structure data
usually comes in a familiar form that looks something like this, in rows and
columns. This is called a table dataset.
Market num_bedroomsnum_bathrooms Sq_ft pool (Y/N)
Value
$207,367 4 3 2635 N
$148,224 3 2 1800 Y
$226,897 5 3.5 2844 Y
$122,265 2 1.5 1644 N

The main question to consider during this process is whether your model
will still work when presented with new data. You can test this through
cross validation. This means that you will test your model for data that you
have not yet used. Have some data on hand that you have not used during
training to see how accurate your model is at the end.
You can also use K-fold validation to check the accuracy of your model.
This method is quite easy to use and generally unbiased. It's a good
technique to use if we don't have a lot of data to test with. For K-fold
validation, we will split our data into K-folds, usually between 5 and 10.
Test each fold and see how they performed over all folds when you are done
testing. Usually, the larger your number for k, the less biased your test will
be.
So far we've talked about models that interpret data to find meaning and
patterns. But what kind of data are we going to use? Where do we get our
data from and what does it look like?
Data is the most critical part of machine learning. After all, your model only
learns with data, so it is important that you have relevant and meaningful
data. It came in many shapes and sizes, different structure depending on the
types of data. The more structured the data, the easier it is to work with.
Some data have a very small structure and these data are more difficult to
interpret. Face recognition data can be huge and has little meaning to the
untrained eye.
Structured data is better organized. This is the type of data you are likely to
use when you first start. It will help get your feet wet and you can start
understanding the statistic involved in machine learning. Structure data
usually comes in a familiar form that looks something like this, in rows and
columns. This is called a table dataset.
Prepare the data
So now you have your data, but how do you get it to a point where it is
readable by your model? Data seldom fits directly with our modeling needs.
In order to properly format our data, a round of data cleanup is usually
required first. The data cleaning process is often referred to as data
scrubbing.
We may have data in the form of images or emails. We have to rewrite it so
that it has numerical values that can be interpreted by our algorithms. After
all, our machine learning models are algorithms or mathematical equations,
so the data must have numerical values to be modeled.
You may also have pieces of data that are recorded incorrectly or in the
wrong format. There may be variables you don't need and need to get rid of.
It can be tedious and time consuming, but it is extremely important to have
data that works and can be easily read by your model. It is the least sexy
part of a data scientist.
This is the part of machine learning where you are likely to spend the most
time. As a data scientist, you probably spend about 20% of your time on
data science and the other 80% of your time making sure your data is clean
and ready to be processed by your model. We may combine multiple types
of data and we will need to reformat the recordings to match them. First, in
the case of guided learning, choose the variables that you think are most
important to your model. Choosing irrelevant variables or variables that
don't matter can cause bias and make our model less effective.
A simple example of cleaning or scrubbing data is recoding a response for
gender. On your data you have a column for male / female. Unfortunately,
men and women have no numerical value. But you can easily change this
by making it a binary variable. Assign female = 1 and male = 0. Now you
can find a numerical value for the effect that being a woman has on the
outcome of your model.
We can also combine variables to make it easier to interpret. Let's say you
create a regression model that predicts a person's income based on several
variables. One of the variables is the level of education, which you have
recorded in years. So the possible responses for years of education are 1, 2,
3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16. These are many separate
categories. You could simplify it by creating groups. For example, you can
rewrite variables 1, 2, 3, 4, 5, 6, 7, 8 = primary_ed and rewrite 9, 10, 11, 12
= secondary_ed and rewrite 13, 14, 15, 16 = tertiary education. Instead of
twelve categories, you have three. Respondents have either a basic
education, secondary education, or a level of post-secondary or university
education. This is known as binning data and it can be a good way to clean
up your data if used properly.
When you combine variables to make interpretation easier, you have to
balance more streamlined data with losing important information about
relationships in the data. Note that in this example, by combining these
variables into three groups instead of sixteen, you create bias in your model.
There are many factors that can cause you to clean your data. Even a
misspelling or an extra space somewhere in your data can have a negative
effect on your model.
You may have missing data. To fix this situation, you can replace the
missing values with the mode of the median of that variable. It is possible to
delete data with missing values if there are only a few, but this just means
you have to use less data in your model.
Programming tools
To process your data, you need special programming tools so you can tell the
data what you want them to do. We have already mentioned that machine
learning is a branch of computer science. This is where that comes into play.
In the introduction, we said that the three most common languages for data
science are Python, R and C ++. Choosing the right one depends on your
experience and what you plan to do with your data.
The most common language for data science is python. It was created in
1991 by Guido Van Rossum and it is notable for being easier to read than
other programming languages. It is still being developed and improved. It is
not complicated to learn and is compatible
with the most relevant data types. It also has applications beyond data
manipulation that will be useful in machine learning.
Python has several free packages that you can install that were created to
give you shortcuts to commonly used data science tools. These packages
contain shortcuts to codes often used in machine learning, so you have less
work to do.
Pandas is an indispensable library of tools for data scientists working with
python. This makes it easier to manipulate time series data and table data
series. It shows your data in rows and columns so that it is easier to manage,
the same way you would look at data in Microsoft Excel. It is easy to find
online and free to download. Pandas are useful when looking at datasets in
.CSV format.
Numpy is a useful program to process data faster with python. It works in
the same way as Matlab and it can process matrices and multidimensional
data. It will help you import large data sets more easily.
Scikit-learn is another library of the machine learning function. With Scikit
learn, you have easy access to many of the algorithms we mentioned earlier,
which are often used in machine learning. Algorithms like classification,
regression, clustering, support vector, arbitrary forest and k resources have
shortcuts so much of the grunt coding is done for you.
R is the third option. It is free to use and open source. R can be used for both
data mining and machine learning. It is popular for those new to data science
because of its availability. It can't handle the larger datasets required for
more advanced machine learning operations, but it's not a bad place to start
if you're new to data science and computer programming.
You need a computer to run these programs. Usually a regular laptop or
desktop computer is powerful enough to handle smaller and medium data
sets, especially if you are new to machine learning.
While Graphics Processing Units (GPUs) have been around for some time,
their accessibility has increased in recent years, making data science more
accessible. It is a
breakthrough in data science because the field is no longer limited to labs
with huge computers.
GPUs are known as the power behind video games. This allows a computer
to interpret multiple points at once, which is essential for processing large
amounts of data. With GPUs, we can now do much more with much less
computer hardware. The predecessor, CPU cores, control multiple control
units, allowing information to be processed in one go. Rather than having
multiple control units, the GPU has a much larger web of cores that can all
handle different processes at once. One GPU card can contain nearly 5000
processors. It is a major advance for artificial intelligence and machine
learning. They can help speed up the processing of neural networks.
C and C ++ are other commonly used languages for data analysis. The
advantage of C ++ is that it is a very powerful language. It can process huge
data sets very quickly. Data scientists who use massive data sets often
choose to use C ++ because of its speed and processing power, especially
when working with data sets over a terabyte. C ++ can process one gigabyte
of data in about a second. This makes it especially useful for deep learning
algorithms, 5-10 layer neural network models and huge data sets. This type
of model can be overwhelming for software that is not so fast. If you are
doing more advanced machine learning and you have multiple GPUs, then C
++ may be the language for you. C ++ can do almost anything; it is a very
versatile language.
The downside is that the libraries in C ++ are not as extensive as those in
Python. This means that when you write code for your data and model, you
probably start from scratch. No matter what kind of projects you decide to
do, there will be roadblocks as you write your code. Having a library that
can help you when you get stuck helps you learn and work faster.
Develop models

You must set up in Python or another programming language to learn

machine learning. You create machine learning models by using code to
manipulate the datasets. While this book does not cover machine learning
coding, I will give you a brief overview of some basic libraries and packages
that I recommend you install for machine learning.
Since this is the most widely used language in data science, we will use
Python as an example in this chapter. I also think it is the most practical
language to learn if it is your first language because it is more readable than
other programming languages, and it has a wide range of possibilities
beyond machine learning.
Once you have the latest version of Python installed, there are a few
recommended libraries to install that contain many commands that can be
useful for your machine learning work. All of these can be
easy to find with a quick Google search and they can be downloaded for
free.
The main library for data analysis and machine learning in Python is called
Pandas. It's a fairly popular choice for data sets and will make your encoding
easier and faster, especially if you're still trying to get a feel for things.
Anaconda for python
Another option to get started with Python is to install Anaconda. The great
thing about Anaconda is that it gives you every package for Python, so you
don't have to install the packages one by one while writing the program for
your model. It comes with all the libraries you need, for just about every
other kind of function.
Anaconda is a free and open source program that works in both R and
Python. With Anaconda you have access to various libraries that will help
you with your data science projects. This basically gives you a prepackaged
collection of all python libraries, of which there are over 100 libraries.
One of the main libraries is Spyder and Jupyter. Both are integrated
development environments, meaning they are the window where you will
write your code, but they are more developed than a standard command
window and have options to store and export / import codes.
Most Python users start in a development environment called IDLE. It is
very simple and offers a good format for learning to code in python. When
you install Python on a Windows computer, it is included automatically. If
you have a Linux computer, it is available, but you must install it separately.
IDLE makes those little steps in Python easier because you can save your
scripts and edit them later. It will also help you debug.
To install Anaconda, go to:
docs.anaconda.com/anaconda/install
Scroll down until you see a list of operating systems. Choose your operating
system. It gives you instructions on how to install anaconda on their website
based on your operating system. Then you are ready to mess around in
Python. I highly recommend using one of the free Python beginner tutorials
available on the web. EdX has a free beginner's guide in Python, which is a
great place to start. Also take advantage of forums such as Reddit, where a
large number of frequently asked questions have already been answered in
detail and members always share relevant news from the world of machine
learning.
Algorithms
Once you have your data and the hardware and software to manipulate it,
you need to bring them together. Put your data on your programming
software. Find a free dataset online to work with when you first start.
Kaggle.com is free and has many datasets to choose from in CSV format,
which you can easily work with once you have imported the Pandas library
into your Python.
The best algorithms to start with are linear and logistic regression for
guided learning and clusters of k-means in unsupervised learning. These
will start relatively easily and you can build to other models from there.
Visualization tools
You have your data, and now you have made models using one of the
programming languages, and you have a whole collection of data science
libraries to help you do all this faster. Your computer works well and you
can make models independently.
You may have created models that yield interesting results, but to break it
down into lay terms and communicate your findings to stakeholders, you
need to organize it in a way that is easy to visualize. If you're a data
scientist in a marketing project, you may have created a model that allows
you to categorize customers and predict trends in buying habits. But if you
want to communicate these results to the rest of your marketing team, you
need to find a way to communicate so that even people who
not familiar with data science can understand your results. By splitting your
data into graphs and charts and visually, you can supplement your analytical
skills. Being able to create visualizations of your data is extremely
important when communicating with an audience unfamiliar with data
analysis
A popular toolset for data professionals is Tableau. Such tools are called
data visualization software. Some companies have employees whose entire
job is to take hard-to-read data and present it in a way that is easy to
visualize.
Software such as tableau is very often used by companies that rely on data
to make decisions. Tableau is useful because it is relatively easy to use and
data can be viewed through the platform in real time. You can customize a
dashboard of tools for creating reports and charts with your data. It also
gives you the opportunity to share your results with other people in your
company. Tableau can be used to create graphs and scatter plots of that data
that you have analyzed in your programming language.
More advanced things that are useful
These tools may not be as relevant to you when you are just starting out, but
it can be interesting to talk about some of them and think about what might
be useful later. This book may just be the beginning of your path to
becoming a machine learning expert, so you can refer to this list later if
you're a little more advanced.
You need to keep thinking about managing unstructured data. Usually this
requires more advanced programs because it is more difficult to manage
and manipulate. This type of data often takes the form of something way
too complicated for the human brain to analyze without the help of tools,
but this is the direction machine learning is heading. Using neural networks
to mimic the functions of human thinking, who knows what the future
holds.
The further we get with machine learning, the bigger our data becomes. The
possibilities of machine learning are increasing. The data that will be
important in the future does not have the neat structure that we are used to,
like the kind of data that fits in an Excel sheet.
This type of data also requires more powerful computer hardware and
software to handle the processing of these large amounts of information.
Usually use some kind of cloud computing software to transport the large
amounts of information, as well as a GPU specific for data analysis. This
higher calculation level can help to process multiple moving points at the
same time. The required math is also getting more difficult. Combine
algorithms.
Epilogue
Hopefully, after reading this book, you will have a good understanding of
the basics of machine learning. You have now become acquainted with
various types of popular machine learning models and their uses. We
explored how advanced data scientists use machine learning to make
predictions and the parameters they need to make predictions that are
accurate and reliable when introduced into new data.
The great thing about data science and machine learning is the wide range
of applications. If you go out and gain experience with machine learning,
there is a wide range of jobs and opportunities that work with all kinds of
data. Whether you like the competition and crowds, and you want to use
models that predict the rise and fall of stocks or guess what a customer will
buy next. Or maybe you are interested in medicine and healthcare; you can
apply machine learning to improve cancer diagnosis and gain new insight
into the characteristics of a disease and how it will affect different
individuals. Wherever your interests lie, there is a chance that machine
learning will be implanted to improve what we can already do.
The more the world becomes connected, the more data becomes available.
Almost everyone has some kind of smart device registration and tracking
their user data. Data scientists find more creative ways to learn from and
interpret that data. Machine learning is a way for data scientists to explore
trends beyond the scope of human understanding, meaning that our
predictions will remain more accurate and our data will be more useful.
Computers will only get more powerful and that power will become more
and more accessible, which means that machine learning and data science
are no longer just buzzwords, but commonly used methods to find valuable
information. It is not only large companies that make more use of data and
machine learning; It is becoming easier for even smaller companies to
include big data in their decision-making processes.
Now that you know the basic theory of machine learning, it's time to move
on and find ways to apply and practice the knowledge. If you're interested
in being a data scientist specializing in machine learning, this book is just
the beginning of the process. I strongly recommend that you commit
yourself to learn a language like Python, R or C ++. The next step is to
become a data scientist and apply these theories of machine learning in
actual models and algorithms. Thanks to the Internet, there are a large
number of free books, videos and tutorials available that will guide you
through the process of learning computer languages. There has never been a
better time to learn how to code and make your own models. This book is
only a small part of a large collection of information available on the
subject. If you take machine learning seriously, this isn't the only book you
read.
You can find whole books that describe the process of specific models.
Neural networks are an area so advanced that you could find entire books
based on that specific type of model alone. It's not a bad idea to pick up a
few statistical study guides so you can refer to them if you have a question.
Be on the lookout for potential data sources you could potentially use and
potential questions that might be interesting to research with statistical
math.
The vacancies alone are sufficient reason to pursue further knowledge in the
field. There is a shortage of experienced data scientists who can apply the
methods and techniques mentioned in this book. This means there is an
opportunity for someone who wants to get their hands dirty and start coding
their own models. There are companies and organizations that are currently
looking for people who can make good use of this information. Keep in
mind that this question will not exist forever. Universities all over the world
are already creating new courses that specifically focus on data science as a
mix of computer science and statistics. This means that the next generation
of data scientists is already on the way.
So start learning now. Find an online tutorial and some free datasets online
and find out how to use regression and classifications, using this book as a
guide. Learn each model one by one. Search the web for examples of
completed models and see if you can replicate the results. Learning
programming languages takes time, so be patient and find new
opportunities to learn and adjust your skills.
Try one of the online communities specific to statistical modeling to cut
your teeth and learn from what other data scientists do. I recommend you
check out Kaggle.com. It is a website that organizes statistical modeling
competitions for aspiring data scientists. Various companies and
organizations place competitions with the supplied datasets. It's a great way
to experiment with different tasks and get data to work. Old contests are
also available online, with a host of accompanying tutorials on youtube.com
and other data science communities for you to learn from. It is probably the
best way for an aspiring data scientist to expand his / her resume and
network with other data scientists.