Supervised Learning –Unsupervised

Learning - Reinforcement

Dr Lokhande
Osmania University
 Supervised machine learning is used
for making predictions from data. In
this approach, the model is provided
with a labeled dataset-which is
known know dataset , The data
includes input-output pairs, meaning
that for each input example, the
corresponding output is already
known. The goal is for the model to
learn to map new inputs to the
correct outputs in other words model
need to find the mapping function to
map the input variable (X) with the
output variable (Y).

Training Phase: The model is trained using a labeled dataset. Each data point contains
features (inputs) and a target variable (desired output).
Learning Process: The model iteratively makes predictions on the training data and
compares them to the actual outputs. The differences between predictions and actual
values are used to update the model's parameters.
 Supervised Learning

Classification problems
Assigns an input to a discrete
category or class.
Different algorithm for it are Regression problems
Predicts a continuous numerical value based on the input
• Support Vector Machine (SVM): features
Different algorithm for it are
• k-Nearest Neighbours (k-NN):
• Naive Bayes: • Polynomial Regression
• Support Vector Regression (SVR)
• Decision Tree Regression
• Random Forest Regression
 •The final output after
processing might
categorize or
summarize the data in
a new way based on
the discovered
patterns. In the image,
it seems that similar
animals are grouped
together as an
example of output.

•This refers to the specific unsupervised

learning technique applied to the data.
•After applying an algorithm, this step
Common algorithms include clustering,
• It involves examining the raw data to identify involves further refining or
principal component analysis (PCA), or
patterns, outliers, and the structure of data without transforming the data based on the
association rule learning.
the guidance of a known outcome or target variable. output from the algorithm. It may
•These algorithms are designed to
•This might include tasks like understanding the include tasks like dimensionality
discover inherent patterns or groupings
various features of the data, such as the shapes, reduction, noise filtering, or
within the data. For example, clustering
transforming the data into a more
sizes, or types of animals in the image. might be used to group similar types of
useful format for visualization or
animals together based on their features.
further analysis.
 Unsupervised Learning

Clustering: Association:

Clustering is a method for grouping a set of Association rule learning is a rule-based machine learning
objects so that those in the same group (or approach for identifying interesting relationships between
cluster) are more similar than those in other variables in huge datasets. For example, if many customers buy
groups. bread and milk together, an association rule like {bread} →
Clustering Types:- {milk} can be established
1.Hierarchical clustering
2.K-means clustering Some common association rule learning algorithms include:
3.Principal Component Analysis •Apriori Algorithm
4.Singular Value Decomposition •Eclat Algorithm
5.Independent Component Analysis •FP-Growth Algorithm
Eclat stands for Equivalence Class Transformation.
FP in FP-Growth stands for Frequent Pattern.
Common association rule learning
algorithms
u Apriori Algorithm: Identifies frequent itemsets and generates association
rules using a bottom-up approach with a candidate generation process.

u Eclat Algorithm: Equivalence Class Transformation :An efficient depth-first

search algorithm for mining frequent itemsets using vertical data format
intersections.

u FP-Growth Algorithm: Frequent Pattern : Uses a tree-based structure to

efficiently mine frequent itemsets without candidate generation, reducing
database scans.
What is Semi-Supervised Learning?

u Semi-supervised learning combines aspects of supervised and unsupervised

learning.
u It uses a small amount of labeled data with a large amount of unlabeled data
for training.
u It helps models generalize better when labeling data is expensive or time-
consuming.
How Semi-Supervised Learning Works

u 1. Labeled data: Used as a foundation for learning.

u 2. Unlabeled data: Patterns learned from labeled data are applied to predict
labels for unlabeled data.
u 3. Pseudo-labels: These predicted labels help refine the model’s learning.
u 4. Learning: The model gradually improves its accuracy using this
combination.
Example of Semi-Supervised Learning
Imagine a student learning a concept:

u • Supervised: The student is taught by a teacher in class.

u • Unsupervised: The student studies alone without guidance.
u • Semi-Supervised: The student first learns with the teacher, then self-
studies, with occasional teacher feedback.
Advantages of Semi-Supervised Learning

u 1. Cost-effective: Fewer labeled data means reduced labeling costs.

u 2. Improved accuracy: Unlabeled data helps generalize better.
u 3. Efficient learning: Solves the drawbacks of both supervised and
unsupervised learning.
u 4. Simple: Easy to implement by combining supervised and unsupervised
techniques.
Disadvantages of Semi-Supervised
Learning
u 1. Unstable results: The use of pseudo-labels can lead to inconsistent
accuracy.
u 2. Limited applicability: Not suitable for all data types (e.g., network-level
data).
u 3. Lower accuracy: Sometimes less accurate compared to fully supervised
learning.
 Key Concepts in Reinforcement Learning
Reinforcement Learning (RL) 1.Agent: The learner or decision-maker.
2.Environment: The system or setting within which
the agent operates.
3.State: The current situation of the agent within
Reinforcement Learning (RL) is a the environment.
machine learning paradigm in which an 4.Action: All possible moves that the agent can
take.
agent learns to make decisions through 5.Reward: A feedback from the environment that
interactions with its environment. The evaluates the success of an action taken by the
agent.
core notion is to learn from the results of 6.Policy: A strategy that the agent employs to
actions rather than being taught exactly determine its actions based on its current state.
7.Value Function: A function that estimates how
what to do. In contrast to supervised good it is for the agent to be in a given state (i.e.,
learning, which trains a model with the the expected total reward for being in a state and
following a particular policy thereafter).
correct answers (labels), reinforcement 8.Q-function (Action-Value Function): A function
learning indicates success or failure via that estimates how good it is to perform a
particular action in a particular state.
rewards and punishments as the agent
 Aspect Supervised Learning Unsupervised Learning Reinforcement Learning
Uses unlabeled data. No Uses feedback in the form of
Requires labeled data for explicit output or labels rewards or penalties, but not
Data Requirements training (input-output pairs). provided. direct labels.
To predict the output for To discover the underlying To learn a policy of actions to
given inputs based on labeled structure or distribution in maximize cumulative reward
Objective training data. data; find patterns or groups. over time.
Indirect feedback through
No feedback. Algorithms must rewards and penalties after
Direct feedback (correct infer structures or patterns in actions, not necessarily
Feedback output is known). the data. immediate.
Total accumulated reward,
convergence time of the
Accuracy, precision, recall, Silhouette score, Davies- policy, efficiency and
F1-score, mean squared error, Bouldin index, clustering effectiveness of learned
Evaluation Metrics etc. accuracy. behaviors.
Model that can predict or Model or representation that A strategy or policy that
classify new examples based describes or reduces the data optimizes the decision-making
Outcome on learned patterns. without explicit guidance. process based on rewards.
Autonomous vehicles,
Predicting customer churn, Market basket analysis, algorithmic trading, adaptive
image recognition, medical genetic clustering, social websites, personalized
Typical Use Cases diagnosis. network analysis. learning environments.
Customer segmentation,
Image detection, Population feature elicitation, targeted Drive-less cars, self-navigating
Example growth prediction marketing, etc vacuum cleaners, etc
Examples of Machine
Learning
Real-World Applications and Impact
Machine Learning Example

• Google Assistant: Helps you find information, set reminders, and answer
questions, using machine learning to understand your voice commands.

• Online Customer Support: Chatbots or virtual assistants help answer your

questions quickly by learning from common customer queries.

• Product Recommendations: When you shop online, machine learning

suggests products based on your past searches and purchases.