FLIGHT PRICE PREDICTION USING MACHINE LEARNING

Project report submitted in partial fulfilment of the requirement for the degree of Bachelor of
Technology

In

Computer Science and Engineering

By

Yadunandan Sood (181297)

UNDER THE SUPERVISION OF

Dr. Pradeep Kumar Gupta

to

Department of Computer Science & Engineering and Information Technology

Jaypee University of Information Technology, Waknaghat,

173234, Himachal Pradesh

 CERTIFICATE

Candidate’s Declaration

I hereby declare that the work presented in this report entitled “Flight price prediction using machine
learning ” in partial fulfillment of the requirements for the award of the degree of Bachelor of
Technology in Computer Science and Engineering/Information Technology submitted in the

department of Computer Science & Engineering and Information Technology, Jaypee University of
Information Technology Waknaghat is an authentic record of my own work carried out over a period
from August 2021 to May 2022 under the supervision of Dr. Pradeep Kumar Gupta, Associate
Professor Department of Computer Science and Engineering, Jaypee University of Information
Technology, Waknaghat.
The matter embodied in the report has not been submitted for the award of any other degree or diploma.

Yadunandan Sood (181297)

This is to certify that the above statement made by the candidate is true to the best of my knowledge

Dr. Pradeep Kumar Gupta

Associate Professor
Computer Science & Engineering and Information Technology

Jaypee University of Information Technology, Waknaghat,

I
 AKCNOWLEDGEMENT

Firstly, I express my heartiest thanks and gratefulness to almighty God for His divine blessing makes us
possible to complete the project work successfully.

I really grateful and wish my profound my indebtedness to Supervisor Dr Pradeep Kumar Gupta,
Designation, Department of CSE Jaypee University of Information Technology,Wakhnaghat. Deep
Knowledge & keen interest of my supervisor in the field of Computational and Machine Intelligence
to carry out this project. His endless patience, scholarly guidance, continual encouragement, constant
and energetic supervision, constructive criticism, valuable advice, reading many inferior drafts and
correcting them at all stage have made it possible to complete this project.

I would like to express my heartiest gratitude to Dr Pradeep Kumar Gupta, Department of CSE, for
his kind help to finish my project.

I would also generously welcome each one of those individuals who have helped me straight forwardly
or in a roundabout way in making this project a win. In this unique situation, I might want to thank the
various staff individuals, both educating and non-instructing, which have developed their convenient
help and facilitated my undertaking.

Finally, I must acknowledge with due respect the constant support and patients of my parents.

Yadunandan Sood

II
 TABLE OF CONTENTS
Chapters Page No.
Chapter 1 INTRODUCTION 1

1.1 Introduction 1

1.2 Problem Statement 3

1.3 Objectives 4

1.4 Methodology 4

1.5 Organization 5

Chapter 2 LITERATURE SURVEY 7

Chapter 3 SYSTEM DEVELOPMENT 10

3.1 Design and Algorithm 10

3.1.1 Exploratory Data Analysis 10

3.1.2 Data Preprocessing: 13

3.1.3 Feature Selection 16

3.1.4 Random Forest Algorithm 20

3.1.5 HTML and CSS 25

3.1.6 Bootstrap 27

3.1.7 Flask Framework 28

3.1.8 Heroku 30

3.2 Model Development 31

3.2.1 Dataset 31

3.2.2 Computational 33

3.2.3 Experimental 35

3.2.4 Mathematical 37

Chapter 4 PERFORMANCE ANALYSIS 40

4.1 Analysis of System 40

Chapter 5 CONCLUSIONS 47

5.1 Conclusions 47

5.2 Future Scope 48

REFERENCES 49

III
 LIST OF ABBREVIATIONS

 EDA: Exploratary Data Analysis

 C: Celsius
 F: Fahrenheit
 ML: Machine Learning
 i.e: that is
 RAM: Random Access Memory
 CPU: Central Processing Unit
 GPU: Graphics Processing Unit
 KNN: K-Nearest Neighbours
 INR: Indian Rupees

IV
 LIST OF FIGURES

Fig 1.1: Depiction of decision tree…………………………………………............2

Fig 1.2: Random Forest…………………………………………………………….3

Fig 3.1: System Design…………………………………………………………….10

Fig 3.2: Example of info function………………………………………………….11

Fig 3.3: Example of describe function……………………………………………..12

Fig 3.4: Scatter plot………………………………………………………………..12

Fig 3.5: Box plots………………………………………………………………….13

Fig 3.6: Histograms………………………………………………………………..13

Fig 3.7: Illustrations of missing values……………………………………………15

Fig 3.8: Split ratio…………………………………………………………………16

Fig 3.9: Types of features…………………………………………………………18

Fig 3.10: Feature vector…………………………………………………………...18

Fig 3.11: Feature importance……………………………………………………...19

Fig 3.12: Heatmap………………………………………………………………....20

Fig 3.13: Illustration of random forest…………………………………………….22

Fig 3.14: Bagging process…………………………………………………………23

Fig 3.15: Basic HTML code………………………….……………………………26

Fig 3.16: Description of target variable……………………………………………32

Fig 3.17: Linear regression………………………………………………………...35

Fig 3.18: Decision tree………………………………………………………..……36

Fig 3.19: Depiction of KNN…………………………………………………….....37

Fig 3.20: Journey_Month vs Price………………………………………………....38

Fig 3.21: Airline vs price………………………………………………………..…39

Fig 4.1: Original dataset……………………………………………………………41

Fig 4.2: Dep_time converted to single units……………………………………….41

Fig 4.3: Arrival_time converted to single units…………………………………….41

Fig 4.4: Airline vs price…………………………………………………………….42

V
Fig 4.5: One hot encoding performed on airline field………………………….......42

Fig 4.6: Source vs Price……………………………………………………………42

Fig 4.7: Label encoding on total_stops field……………………………………….43

Fig 4.8: Airline vs average price…………………………………………………...43

Fig 4.9: Scatter plot of duration vs average price…………………………………..43

Fig 4.10: Original test set………………………………………………………......44

Fig 4.11: Graph of feature importance……………………………………………..44

Fig 4.12: Scatter plot between the test and predicted values……………………….45

Fig 4.13: Accuracy before hyperparameter tuning…………………………………45

Fig 4.14: Accuracy after hyperparameter tuning…………………………………...45

Fig 4.15: Flask App…………………………………................................................46

Fig 4.16: Final Output…………………………………............................................46

VI
 LIST OF TABLES

Table 3.1: Types of converter variable…………………………………………….29

Table 3.2: Dataset information…………………………………………………….31-32

Table 3.3: CPU specifications……………………………………………………..33

Table 3.4: GPU specifications……………………………………………………..34

Table 3.5: Dataset stats…………………………………………………………….38

Table 4.1: Models and precision…………………………………………………...40-41

VII
 ABSTRACT

Machine learning is a field of study that works with computer algorithms and data. It basically builds a
tool on sample data and then predicts or take decisions about the same data without the need of specific
programming. ML has its use in various industries including aviation industry where it is used to predict
flight prices.

In today’s world, flights have become a common mode of transportation for humans. It is faster way to
reach destinations, thus saving a lot of time. But also not everyone can afford a plane ticket and fly high
in the sky. The ticket prices for a plane journey are very high when compared to that of a train or a bus.
But irrespective of that, nowadays the number of people using flights have increased massively. Thus it
becomes very hectic to maintain the prices of tickets with changing conditions. The airline companies
use various complex techniques to predict the various flight prices using various factors present at that
time. These factors commonly include market related issues, financial issues and various social issues.
So it becomes difficult for them to do so. Also the consumer has no such tool in hand to have an idea
about the ticket prices, that could very helpful.

This project is designed and developed keeping in mind this problem and develops an algorithm that
predicts various flight prices keeping in mind various factors that affects them. This can help the airline
companies to check what prices they could maintain. Also it would help the customer to predict future
flight ticket prices, which would help in planning the trip accordingly.

VIII
Chapter 1 INTRODUCTION

1.1 Introduction
Machine learning is the part of artificial intelligence that provides a computer the ability to
automatically learn with experience without the need of external programming. It is one of
the hot ,most in demand research topic in computer science engineering. It can provide
intelligence to the machines with the help of various tools and techniques. ML uses programs
that can access data and learn for themselves using the same. In today’s world machine
learning has become an important part of everything we are a part of. It helps various
enterprises for the development of new products and also helps to have an idea about the
customer trends. Many MNC’s have made ML an important part of their organisation.

The learning process for an ML algorithm begins and ends with data. Therefore, data is the
most important part of a model. This data has an effect on future predictions and decisions.
The primary objective of ML is that a computer system learns automatically without human
assistance.

Machine Learning is categorized mainly into 4 approaches based on how an algorithm goes
through the process of learning.

 Supervised machine learning: Analysis of labelled data and then training on the
same to predict future events. This system can provide output for new inputs if
sufficient training is done.
 Unsupervised machine learning: The information used in training is not labelled
and neither classified. This system is used to study different patterns and connections
in data.
 Semi-supervised machine learning: Falls in between the above two as it uses small
amount of labelled data and larger amount of the unlabelled data.
 Reinforcement machine learning: In this system the program learns to behave in a
particular environment by performing actions and then seeing the results.

Decision tree is a powerful tool used for classification and prediction. It resembles the shape
of a flowchart. They classify instances from the root node towards the leaf node. A condition
is specified at each node(internal node) based on which the tree further divides into edges.

1
The end of the branch, i.e a branch where no further splits occur is called the leaf or the
decision.

Root Nodes: First node at the start of the tree. Division starts here.

Decision Nodes: Nodes that we get after splitting of root nodes.

Sub-tree: It is basically a sub part of the whole tree.

Pruning: Cutting of some nodes to stop overfitting.

Figure 1.1: Depiction of decision tree

Random forest algorithm is a popular model in machine learning that comes under the
supervised learning approach. This model can be used for both regression and classification
problems. This model takes together a number of classification models or commonly known
as classifiers to solve a problem and hence improve the accuracy of the same. Random forests
contains a number of decision trees and predicts by taking the collective output of them ,i.e
it takes the mean or average of the outputs from various trees. Also it can be inferred that by
increasing the number of trees, increases accuracy. A main fundamental behind this classifier
is that a large number of uncorrelated trees operating as a single unit have to be taken. This
sense of least relationship between the trees would help to maximize the results. Example
like in investments with low relatedness (like bonds and stocks) , form a portfolio that is
much greater than individual sum of its parts.

2
 Figure 1.2: Random forest

1.2 Problem Statement

In today’s world flight companies try to change flight prices according to their needs and
present condition. In this way they can maximize their profits. Regression based solutions to
this problem are common but they come with their own disadvantage. With the
technological advancements and the load increase on these flight companies , regression
does not give the desired results and offers low accuracy. Customers on the other hand do
not have a tool to have an idea about the ticket prices that could help them in planning
accordingly. Also many customers don’t know the right time to book a ticket and sometimes
have to pay more for a ticket. Consider an example, the least expensive ticket will change
its value over a period and its value may be high or low depending on the time like summers
,winters etc or according to day ,night or evening. This happens because the underlying goal
of the plane carrier is to maximize its income. But on the other side the purchaser try to get
a ticket at a lower price so book them days before the actual takeoff. But sometimes the price
of a particular ticket does not fluctuate and he consumer ends up paying more for the same.

So random forest machine learning model would be used for the same problem to get the
prices of flight tickets..This was a brief into the problem which will be discussed below in
the report and a possible solution to this problem will be provided.

3
1.3 Objectives

This project aims at developing a model that would predict the prices of the flight from the
starting point to the point of destination and should be efficient in working. The price
prediction would depend on various factors such as distance, flight time, number of stops
etc. These factors help create a pattern to decide the price of flight ,and the machine learning
models get trained on this pattern to make predictions in future. A model would be created
by applying machine learning algorithms to the collected data or the historical data related
to flights. This system would give people the idea about the trends that prices follow and
provide a predicted value.

Nowadays the airline corporations are using complex methods to get airfare prices in a
changing fashion. These strategies take into account several commercial, social and
marketing factors. Due to the high variedness of the pricing models applied by the airlines,
it is very difficult for a customer to buy a ticket in the lowest price possible. The model
should be able to give an insight on the fare prices for the desired flight. Also it should be
able to predict the prices for maximum number of airline companies, i.e it should be able to
adopt to new changes or better say to new data. The accuracy offered by the model should
be high so that it could be trusted and customers don’t lose their interests in knowing the
prices beforehand. Also the model should not have any unnecessary lags or stops in its
working as the scenarios its dealing with does not allow any slip ups.

1.4 Methodology

The main methodology of this project revolves around using the random forest algorithm for
classification and prediction. First and foremost we begin our project with collecting the
required data and applying the exploratory data analysis. This is nothing but initial
investigations on data to look for any patterns or to check if there is something wrong with
it with the help of statistics and graphical ways. After this we move on to data preprocessing.
Data preprocessing is the step where data is changed or encoded so that the machine could
easily understand it. It also involves the process of dealing with missing values, duplicate
values or inconsistent values. Then we move onto feature selection. A feature in ML is an
attribute or characteristic under observation. For example for a flight power , fuel capacity,

4
number of stops, weather conditions can be considered as features. So in feature selection
we pick out the best features or the features which when applied to the model gives results
with high accuracy. Feature selection is an important of supervised learning because if we
do not go ahead with the right features the results may vary a lot that would simply nullify
the purpose of this flight price predictor.

The next step from here is to decide which model to apply to bear fruitful results. We
experimented with models like linear regression, decision trees, multilayer perceptron etc.
After experimenting with all these algorithms we move forward to give our approach for the
problem at hand using random forests classifier.

1.5 Organization

In this part we will explain the organization of the chapter wise formatting of this whole
report. In chapter one we have already seen the Introduction, Problem Statement. After that
we saw the objectives and the methodology of the report. Now we stand in the Organization
section that would explain the chapters and topics involved further down in this report.

Chapter 2 will be presenting a brief overview of the literature survey process. This will
include some of the resources we used in the past to gain necessary knowledge in view of
this project and other work we came across through the community that set us forward to
reach to this stage where we were able to complete this project and develop a basework to
solve these type of problems in the future.Then will follow Chapter 3 where we would be
discussing about how our model is designed and developed. First we would discuss abouth
the dataset that we used in making this report and see some basic features.We will discuss
about the basic framework which help us lead to random forest algorithm that include
decision tree and feature selection. We would take a look at the various models we applied
to draw comparisons with the random forest and their respective accuracies. We have
explained them a bit to give the reader a better understanding of them. After that we have
also listed some of the mathematical formulas used in this report.

In chapter 4 we have shown how our system performed and also compared them with the
other systems used earlier like linear regression ,decision tree etc. Then we begin to
showcase the outputs at various stages of our project work that include various graphs and
figures. This would conclude with the final output of our model.

5
Finally, in chapter 5 we started the conclusion part and explained about what we managed
to achieve in the project and how this model was the perfect way for us to learn and explore
in this field. After that we proposed some future plans that can be achieved through this
project and made an extension to it.

6
Chapter 2 LITERATURE SURVEY

Travelling by the mode of airplanes from one place to another is perhaps the easiest and
most convenient ways but selecting the right type of flight with minimum ticket expense and
the perfect route is a cumbersome process for a customer. Hence sophisticated Artificial
Intelligence technique called Machine Learning provides a suitable and convenient platform
to the users for selecting the right type of flight with minimum price and suitable route. Flight
ticket price prediction is a challenging task as the various factors/variables determining the
cost can change constantly over time.[1]

Flight prices can vary drastically depending on whether the day was a festival, weekday,
weekend, peak season etc. Traditional factor like the total distance between source and
destination is not the sole factor which determines the price but has a broader aspect
depending on a number of factors. Generally the flights that have departure during weekends
have a higher price than those which departure during weekdays [5].Also flight prices during
festive season are generally lower. The route and number of stops also play a major role in
determining the price of the flights. Features like total number of stops, Duration hours,
journey day are some of the most important features for the prediction purpose as the price
of the flight tickets is highly dependent on these types of features. To find the most important
features correlation heatmap is used which gives the correlation values and higher the
correlation value between two features higher are the chances of one of them being removed.

Various Machine Learning algorithms have been used to predict the flight ticket price
whether on international level or national level. A number of datasets have also been used
for the same although the main essence of prediction is the same for all which deals with
proper preprocessing and feature selection. Various algorithms like Linear Regression,
Decision tree, Support Vector Machines (SVMs) and Neural network are commonly used
for building this model. Among all the algorithms mentioned above, Random forest gives
the best accuracy. SVMs fail to give a desired accuracy on regression analysis however it
can be used for classifying the price as high or low.[3]

The following algorithms were tested for accuracy among which Decision tress proved to be
the best:

1) Linear Regression

2) Decision tree Regressor

7
3) Random forest

Linear Regression however does not prove to be a suitable algorithm and only gives an
accuracy of 61 percent. As the numbers of features are more in number and complex in
nature, linear regression fails to give the expected accuracy. Decision tree however gives an
accuracy of 72 percent which is quite good for a dataset of this size.[2]

Random forest is the most suitable algorithm as it uses multiple decision trees to get a final
output and thus results in maximum accuracy among the above mentioned algorithms. It
gives an accuracy of 79 percent and along with hyper-parameter tuning the accuracy jumps
to almost 81 percent. Random forest has almost equivalent parameters as a decision tree.

In order to find the best features for fitting the model, feature selection is very important as
it investigates the degree of impact of each feature on prediction output. Correlation between
the features determines how closely the features are related to one another. One of the two
highly correlated features are removed so as to decrease the size of features and decrease the
complexity in fitting the model.[2]

The above table represents the various ML models along with their accuracies and execution
time. This table was one of the motivating factors to go with Random forest due to its high
accuracy.

8
Various evaluation metrics have been used to get the accuracy and model performance which
helps us to improve the model by accordingly adjusting the parameters or algorithms.
Depending on these metrics we find a suitable algorithm.[4]

The various evaluation metrics used are:

1) MAE (mean absolute error)

2) MSE (mean square error)

3) RMSE (root mean square error)

9
Chapter 3 SYSTEM DEVELOPMENT

3.1 Design and Algorithm

The main algorithm used in this report is random forest. But before explaining that we would
explain little bits about the other stages of our model too, i.e exploratory data analysis , data
preprocessing and feature selection. Also we would try and develop a comparison between
decision trees and random forest and see why the latter is more efficient. Also we have
decided to make a webapp for the same so that the model becomes more expressive and easy
to use. We used HTML,CSS Bootstrap and Flask to create the same. These concepts will be
elaborated further down in the report.

Figure 3.1: System Design

3.1.1 Exploratory Data Analysis

Exploratory data analysis is the process of critical inspection on data so that we can have an
insight on the patterns of dataset, to check for any anomalies with the help of graphs and
statistical analysis. It is a very great step to understand the dataset and gather as many
information from it. It helps to determine how to manilpulate data to get the answers we
need. EDA was originally developed by the mathematician John Turkey in the 1970’s.

10
Some tools and techniques that can be used for the same are :

 Clustering and dimension reduction techniques: These help create graphical

representations of data high in dimensionality.
 Univariate and bivariate analysis: These help to find a relation between the input
variable and the target variable.
 K-means clustering: An unsupervised learning model where K groups are made
based on distance from each group’s centroid.
 Prediction models like regression, that use stats and data for outcomes.

We can use different functions available in the python pandas library to begin our EDA.

The “.head()” and the “.tail()” functions show the first five and the last five rows from the
dataset respectively. It is a good practice to do so as it given an idea about the pattern of data.

We can also know abouth the total of rows and columns in any dataset using the “.shape”
function in pandas. We can also use “.info()” function which help to know the columns and
their data types, and also helps to find out whether any column contains null values or not.
The “describe()” is like the most handy one among all these. It returns various information
about the data like count, mean, minimum and maximum values ,standard deviation etc.

Figure 3.2: Example of Info function

11
 Figure 3.3: Example of Describe function

Data visualization also comes under the EDA as it helps the user to get a graphical insight
into the data. This is basically divided into three categories:

 Univariate analysis: This displays all the observations in data around a single data
variable. Ex: line plot, histogram etc
 Bivariate analysis: Reveals relationship between two data variables. Ex: heat maps,
box plots etc
 Multivariate analysis: reveals relationship between more than two sets of variables.
Ex: violin plots, box plots, histograms etc.

Figure 3.4: Scatter plot

12
 Figure 3.5: Box plots

Figure 3.6: Histograms

3.1.2 Data Preprocessing:

Whenever we think of datasets , a large number of rows and columns pop into our heads.
But this is not the case always because data can be in structured ,unstructured, tables, images,
audio files, videos etc. Now as we all know our computers do not understand anything other
than the language of 1’s and 0’s. So it will be utterly make no sense to pass set of videos or
images to our model. Also a dataset may contain missing values, duplicates that are of no
use to the model or can cause errors. So here comes the part of preprocessing. EDA and
preprocessing are closely related terms and sometimes used in the same way.

13
In any ML procedure, data preprocessing is the step where our data is transformed or in
complex terms encoded ,to bring it to a state so that the computer program can understand
it. The columns or the attributes of the data can be consumed or understood by the algorithm
after preprocessing. Also it helps in cleaning the data i.e getting rid of the various null or
missing values and it helps in increasing the accuracy and efficiency of the model.

This whole process can be divided into the following sub categories:

 Getting the dataset: The collection of data for a particular use is called the dataset.
As a machine learning model completely depends on data so we have to make sure
we acquire the best possible data out there.
 Importing libraries: Some python libraries that we need in our project foe some
specific jobs have to be integrated. This is called importing libraries. Mainly there is
use of three python libraries

Numpy: Used for mathematical calculations in the code.

Pandas: Most common and useful library. Used for data manipulation and analysis
of dataset.

Matplotlib: Used for plotting various charts using python. Works with the help of a
sub library pyplot.

 Importing datasets: The step where we import required datasets for the model

 Handling missing/inconsistent/duplicate data: It is a common thing to have

missing values in a dataset. The cause to it can be anything ,whether it was by default
in the dataset of happened during data collection. But it is our responsibility to handle
these missing values.

1. Eliminate the rows having missing data: It is a simple and effective

strategy. This could also fail if many objects have missing values.

14
 Sometimes the feature has to be eliminated if its has mostly missing
values.
2. Estimation of the missing values: If only some values are missing then
various interpolation methods can be used to fill up these missing values.
But in general dealing with these missing values is done by filling them
with median ,mode or mean from the respective column.

Figure 3.7: Illustration of missing values

3. Duplicate values: Deduplication is an often used term that refers to the

process of dealing with duplicate values. Duplication of values is also
another problem with data. In real world example it can happen when a
person fills in his student details twice when asked by the company. The
reason behind removing of duplicates is that it should not give a
particular data object an edge over the other or in simple words a bias
while learning algorithms.

4. Inconsistent values: Like duplicate values, finding inconsistent data in a

dataset is highly common. For example the flight number in some rows
could be sometimes written under the passenger phone number. The
cause to this could simply be a human error or the information was not
read correctly at the time of filling the entries.

15
  Encoding categorical data: Categorical data is that whose values are taken from a
pre defined set of values. A simple example can be months of a year :{January,
February, March………}
as its values is always taken form this set. Now as an algorithm works on number
and maths it is necessary to convert these into suitable format and can be done using
various methods like LabelEncoder(), OneHotEncoder etc.

 Splitting of data set into training and test set: If we train our model on a dataset
but when it comes to testing we use a completely different dataset we would create
difficulties for the model. Doing this would also decrease the performance of our
model. So we make a model that does well on the training and on the test set. So the
test set can be defined as a subset to the training set to test our model on. Validation
data is also another term used in ML. We use this data to improve hyperparameter
tuning. The model does not learn on this validation data set.

Split Ratio- It is largely dependent on the model we our building and on the dataset.
If a lot of training is required then we use a larger portion of the data for training
purposes ,example image data as it contains millions of features.

Figure 3.8: Split ratio

3.1.3 Feature Selection

Before jumping onto feature selection first lets take a look at what do we mean by the term
features. A dataset is a collection of various objects that are defines by a number of features,
which simply tell us about the different characteristics of our data. Features are also called
attributes, fields or variables. Like for example the mass, height of a person can be features
of a dataset regarding lifestyles of various people. A feature is an individual characteristic
of a phenomenon. We have to choose unbiased and independent features for ML models so
that we can have accurate regression and classification. Features are predominantly

16
numerical but in some cases they can be graphical or strings. For example, for a scooter,
colour, weight, mileage cane be observed as features. In a speech recognition system
noise,sounds,power can be used a features. In algorithms using spam detection certain email
headers, structures, frequency of words can be used as features. Also in character
recognition, the length of a particular letter, the boldness of the letters, the spaces between
words are used as or can be used as features.

Now features can be:

1. Categorical: These are the features whose values are chosen from a predefined
set of values. Example dates in a month {1,2,3,4,5,……..} is a categorical value
as it is chosen always from this set. The Boolean set is also an example of
categorical data.
i. Nominal data: These are the values that don’t have any specific order.
Example a new scooter comes in three color- green ,blue, red.
ii. Ordinal data: These are the values that have a natural order within them
but the difference on scale is not defined. Example when we go for
buying clothes the sizes have a natural order small<medium<large but
this does not ensure that the difference among them would also be the
same.

2. Numerical: These are the features whose values are either discrete or
continuous. They represented in the form of numbers and mostly
posses their properties. Example speed of a plane or the number of steps we walk in
a day.

i. Interval data: This contains a defined unit of measurement and the

difference between the data values is meaningful. Example include
dates of calender ,temperature in C or F.
ii. Ratio : This also contains defined unit of measurement but both
differences and the ratio are meaningful. Examples Age, mass ,
height etc.

17
 Figure 3.9: Types of features

Features can also be represented as feature vectors. In machine learning a feature vector is
an n dimensional vector that represent some object. Many ML models require numerical
equivalent of an object to perform statistical processing. Example in order to represent
images we need to have numerical representations and that is provided y feature vector
,which contain the pixels of on image. Feature vectors usually combine with a set of weights
and use dot product to produce a linear predictor function in order to determine a score for
making a prediction. The space vector associated with these are called feature space. Also
we can develop new features or higher level features from the already available features and
adding it to the feature vector. This is called feature construction.

Figure 3.10: Feature vector

18
In many scenarios using all the features of a dataset is not an ideal approach as some features
do not contribute in increasing the efficiency of a program and also the data size would
become very large. The three main goals of feature selection are to

 Reduce the cost of computation

 To improve the overall accuracy of the model
 Producing a more understandable model

Feature selection can be performed both before or after training ,it can be performed
manually or by automated methods. Below are some methods.

Correlation plot: It is one of the manual techniques to perform feature selection. This creates
a visualization that plots the correlation measure for every feature in the data. We can then
observe which features are closely related and therefore can remove some of those , and also
some variables might have a low correlation with the output variable , so it is recommended
to remove some of those.

Below mentioned two techniques are also used by us in development of our model.

Feature importance: It is a technique that gives a score value to input features based on the
fact that how useful they are in predicting a target variable. This plays an important role in
selecting the best features for our model and increasing the efficiency as it acts as a good
indicator of the features that have an impact on the model and we can remove those that do
not have a significant one.

Figure 3.11: Feature importance

19
Heat maps: Heat maps are good technique to find the correlation between the dependent and
the independent variables. These are a form of matrix plots. We can gain useful insights from
this method.

Figure 3.12: Heat map

3.1.4 Random Forest Algorithm

Before studying random forests we should know about the building blocks and that is
decision trees. Decision tree forms a tree like structure and supports decision. In simple
language decision trees are an advanced version of if-else statement. A decision tree consists
of three components –

Root nodes- Nodes present at start of tree.

Decision nodes- Nodes after splitting.

Leaf Nodes – Nodes where no further split happens.

A decision tree simply takes the training data and then splits it till we reach the point of leaf
nodes. The leaf node cannot be further splitted. The nodes of a tree are the features that are
used for predicting the results of the model. Decision trees are often used in performance
research and project management. If, in fact, decisions have to be made online without
having to reconsider under incomplete information, the decision tree should be compared to
the opportunity model as the best selection model or the online selection algorithm. Making
a decision tree involves the process where we have to choose the features to act on and the
condition on which a split would occur.

20
Splitting cost: It differs in problems of regression and classification. But the common thing
in both the cases is that the cost functions always try to find branches of the tree with groups
that have similar responses. Below are given the general formula used for splitting.

Now the next question arises is that how do we know when to stop splitting in a decision
tree. When a dataset contains large number of features , it would result in larger number of
splits and can lead to overfitting. So there are many ways to stop this from happening by
hyperparameter tuning, example setting the max depth of the tree , or setting the value after
which a split would occur.

Entropy and information gain are the two main bricks that build a decision tree.

Entropy: In simple terms it refers to the measure of noise or randomness in the dataset. In a
decision tree mostly the output is either yes or no. So there has to be a clear distinction.
Example if a certain decision gets 5 upvotes and 4 downvotes thsis would to lead to
disorderness as there are almost equal number of votes for both the ways it becomes difficult
to decide. It measures the impurity of a node i.e, the randomness. A pure split occurs if we
get all the answers of a split from the positive class or from the negative class. When the
purity of a leaf node reaches 100 percent we make that node as leaf node. Higher the entropy,
higher will be the impurity. By means of entropy we can only know about the impurity of a
particular node and have no idea abouth the entropy change in the parent or any particular
node. So information gain steps in here. The formula for entropy is

E(S) = -p(+) log p(+) – p(-)log p(-)

where p+ refers to probability of positive class

p- refers to probability of negative class

S is the subset of the training data.

Information Gain: Plays an important part to decide whether a node is to be used as a decision
node or the root node. It also measures the reduction in randomness according to a given
feature. It can be called the entropy for a full dataset. The formula for information gain is

21
 Information Gain= E(Y)- E(Y|X)

Now we proceed with the random forest algorithm. As obvious from the name random
forests consists of large number of singular trees that operate as a single unit . Each tree in
the forest makes a class prediction and the class which gathers most of the votes becomes
the final prediction.

Figure 3.13: Simple illustration of random forest

Also we have to keep in mind the basic fundamental behind this is that tress with low
correlation are to be used. A greater number of least correlated trees operating as a single
unit would result in high efficiency. The simple reason behind this is that the trees would be
safe from each others individual errors.

Random forest works on the ensemble technique which simply means combining multiple
models. Ensemble uses two types of methodologies-

 Bagging: Also known as bootstrap aggregation. Decision trees are very sensitive on
the training data, i.e even the small changes to it can have a big impact on the tree

22
 structures. It makes a new training subset from sample training data with the help of
replacement and the final prediction is based on majority voting. Therefore each
model is generated from the Bootstrap samples provided by the data via replacement
and this is called row sampling. Now this is called bootstrap, i.e the process of row
sampling with replacement. Example if we have a training data with n features and
goes something like this [5,6,7,8,9], we can give one of the trees the following set
[5,5,6,7,7] i.e values from the same set but provided randomly. Now we have
separate models generating their own results and the final outcome is based on the
maximum voting. The step where all the results are combined and the output is
generated is called aggregation.

Figure 3.14: Bagging process

 Boosting: It simply combines the weak trainers into strong trainers by creating
sequential models such the final model results is higher efficiency. We are not going
to elaborate on this as this technique does not find its use in random forest.

Feature randomness- When we talk about a normal decision tree, the time when we have to
do a split at the node, we have to consider all the features but ultimately have to pick one
that would give us the most amount of separation between the right node and the left node.
But in a random forest , it only has random subset of features to its disposal which forces
even more separation among the individual trees and leads to less correlation among them
and therefore more efficiency.

23
Steps involved in the random forest algorithm:

i. First we have to take k number of random records from a set that has a total of n
records.
ii. Now for each sample data a decision tree is made.
iii. In the next step, the individual trees would generate an output.
iv. The final output is based on average in the case of regression and and majority votes
for the case of classification.

So ultimately in random forest we are left with trees that are trained on various sets of data
(bagging) but also use completely random features to make decisions.

Important characteristic of random forests-

 Variety- not all attributes/variables/capabilities are considered even as making

an character tree, every tree is distinct.
 Proof against the curse of dimensionality- since every tree does now not bear in
mind all the features, the characteristic area is reduced.
 Parallelization-each tree is created independently out of various statistics and
attributes. this means that we will make full use of the CPU to build random forests.
 The brake up in train and test data- In a random forest we don’t should segregate
the facts for train and test set data as there'll usually be 30% of
the information which is not visible by using the selection tree.
 Stability -balance arises because the end result is based totally on majority vote
casting/ averaging.

So now lets us try to draw a comparison between the two and see as to why random forest
just outperforms the decision tree and why we chose it to employ in our project. A decision
tree might be simple to understand and employ but sometimes it can lead to over complex
trees that do not come to good use. A decision tree many times suffer from the problem of
overfitting as the user sometimes does not controls it, but in the case of random forest the
overfitting is handled quite well as model is trained on subsets of data and final result
comes from average or voting. A random forest suffers via slow computational power as it
works on many trees simultaneously and hence the decision tree comes up first in this
aspect. Also they go through higher training periods as compared to decision trees. In
random forest, the data set is not set up on any rules or formuals , i.e random features are

24
selected on their own, but in the case of a decision tree some predefined rules are set to do
the prediction. The main disadvantage of a decision tree is that the slightest of change in
the data can cause disruption to the tree structure.Decision tree novices create
biased trees if some instructions dominate.
Itis consequently advocated to balance the information set prior to fitting with
the decision tree.In an ideal world we would want to reduce the variance and bias based
error and random forests are just perfect for that. They offer more resilience than decision
trees. Based on all these factors we can say that random forest is one of the best and highly
efficient techniques and why we chose to go along with it in our model too.

3.1.5 HTML and CSS

HTML is short for Hyper Text Markup Language and is used to create webpages and
webapps. Tim Berners-Lee is called father of HTML. In 1991 Tim , proposed the first
available description of HTML called HTML tags. Now let us divide into sections.

Hyper Text: This in simple words means text within text. A link is hiddenly stored within
the text. Whenever someone clicks on a link that takes you to a new webpage, this means
you have interacted with a hypertext. It is way to link one webpage to another.

Markup language: It can be defined as a language used to apply layout and formatting to
text document. It makes the text more visually appealing and interactive. Text can be
converted to links , tables and images etc.

Webpage: It is usually written in HTML and then a web browser translates it. URL is used
to identify a web page. Webpages are of two types static or dynamic . But with HTML we
can create static web pages.

HTML is made up of different HTML tags and content is different for different tags.
Below given is the figure of a simple webpage code written in HTML.

25
 Figure 3.15: Basic Html Code

<!DOCTYPE> : It simply tells the web browser about the version of HTML.

<html>: It tells the browser that the current document is of type HTML. All things written
between html tag describes the web document. It also acts as a container for the other
elements except <!DOCTYPE>.

<head>: First tag inside the html tag. Contains metadata. Should be closed before the body
tag.

<title>: It is optional. It basically is the name displayed on the tab of the web browser
window. Should be placed in the head tag and closed immediately.

<body>: Defines the content visible to the user.

<h1>: Top heading level

<p>: Stands for the paragraph in the webpage.

CSS: Stands for cascading style sheets and allows you to create beautiful looking web
pages. With the use of CSS you control how elements of HTML will look in the browser
according to your design. CSS can perform basic actions like change of color and size etc,
but can also do things like animation. In CSS we have to define rules containing styles and
that should be applied to particular elements in the HTML. There are three ways to add
CSS to a document.

 Inline CSS: applies style to a single element. Add the <style> attribute to the
required element.
 Internal CSS: <style> has to be added inside the <head> tag.

26
  External CSS: In this the html document should have a link or a reference to the
external CSS file. The reference has to be provided inside the <link> tag inside the
head section. The external document should be saved with a .css extension. Also it
should be free of any HTML tags.

<link rel=”stylesheet” href=”styles.css”>

Below is a simple example.

It opens with a selector, in this case h1 which tells the doc that we have to style h1
elements. Followed by curly braces and then property-value pairs and ends with the colon.
We can only apply values to a property that are allowed by CSS.

3.1.6 Bootstrap

It is a open source and free CSS framework, used to create responsive, mobile friendly
websites. It is a framework for front end development. It is much easier and faster than
other available platforms. It comprises of CSS and HTML designed templates to do
various stuff like creating forms, tables , carousels, images , text etc. It also supports js
plug ins. It was made at twitter and offered to help a solution to the inconsistencies
between the different frameworks used for front end development.

Bootstrap styles the html elements and can make the webpage uniform. It has predefined
components like tables , forms, cards etc which can be used directly in the project and
customized with the help of CSS classes, according to the developer or user needs.

The most awesome feature of Bootstrap is the layout management or the grid system that
simply brings uniformity to the webpage. The basic is called the ‘container’ and every
other element is placed inside it. Containers can be fluid or fixed. A fluid one is flexible or
responsive and fills the entire width of the webpage. But the fixed one uses fixed widths

27
depending on the screen size as defined by bootstrap. Inside the container bootstrap
implements a row and column structure called CSS Flexbox.

3.1.7 Flask Framework

It is a python written micro web framework. It does not essentially require the particular
tools or technologies and is therefore called micro framework. It does not have any form
validation or a data base abstraction layer where some useful functions could be provided
by the existing third party libraries. But with this flask has its support for extensions that
add app features and looks like if they were written and implemented in it. Extensions exist
for form validation, mappers and various related tools.

It has various components:

Jinja: It is an engine for templates for python. It handles templates.

MarkupSafe: It is the string library for the python language. It marks the strings as safe and
when used with regular strings automatically helps in escaping unmarked strings.

ItsDangerous: It is the data serialization library. Its main use is to store the session of a
Flask application inside of a cookie and prohibits users to tamper with the contents of the
session.

Werkzeug: It is designed for python and is a utility library for Web Server Gateway
Interface apps.

SAMPLE CODE:

From flask import Flask

app1 = Flask(__name__)

@app1.route(“/”)

def hello():
return “<p>Hello</p>”

We started by importing the Flask class. Our WSGI application will be an instance of this
class. After that, we make an instance of the class. The name of the application's module or
package is the first argument. __name__ is a handy shortcut for this that works well in most
cases. Flask needs this to know where to look for resources like static files and templates.

28
The route() decorator is then used to notify Flask which URL should be used for calling our
function. The message we want to show in the user's browser is returned by the function.
Because HTML is the default content type, the browser will render HTML in the string.

Routing: To assist consumers, modern web apps employ sensible URLs. Users are more
likely to like and return to a page if the URL is memorable and can be used to get directly to
the page. Have to use the route decorator to bind any function to the URL. If we write route
(“ / ”) this simply binds with the home page. But if we use something like this route(“/hello”)
then the webpage is bounded with the hello URL.

The variable name> tag can be used to add variables to a URL. The variable name> keyword
argument is subsequently passed to your function. Simple example such as
app1.route(‘/user/<username>’) show that the variable username will be passed to the
function whenever the URL is called. We can also specify the type of argument for the
function with the help of a converter, like this <converter:variable_names>. The various
converters accepted by the route method are given below in the table.

Table 3.1: Types of converter variables

string Accepts any text without a slash

path Same as string but also accepts slashes
uuid Accepts UUID strings
int Accepts positive integers
float Accepts positive floating point values

HTTP methods: When visiting URLs, web applications use various HTTP methods. As
you work with Flask, you should become familiar with the HTTP methods. A route only
responds to GET queries by default. The route() decorator's methods argument can be used
to handle various HTTP methods. If GET method is in use flask automatically supports the
OPTIONS variable to be used by a HTML form. The POST method is basically a tool that
sends the data collected in a form to the server or the URL. Basically it means that the data
we collect in a form is sent to the required URL for the required function when we mention
the POST method in the route decorator.

29
Static Files: If you want to create a dynamic web application then you would need static
files. That’s usually where we find use of the CSS and javascript. When we create a folder
named static in the project it will be available as /static in the application. For generating
URL’s for static files use the following code.

url_for(‘static’ , filename=’style.css’) . The file has to be named as static/style.css .

Rendering templates: It's neither enjoyable or easy to generate HTML from Python
because you have to do all of the HTML escaping yourself to keep the application secure.
As a result, Flask automatically configures the Jinja2 template engine for you.

The render template() method can be used to render a template. All you have to do is supply
the template engine the name of the template and the variables you want to pass as keyword
arguments.

def hello(name=None)

return render_template(‘hello.html’, name=name).

Flask would look into the /templates folder for the required templates used in the program.

Pickle Module: For serialising and de-serializing a Python object structure, the Python
pickle package is utilised. Pickling an object in Python allows it to be saved on disc. Pickle
"serialises" the object before writing it to the file. Pickling is a Python function that converts
a dict, list or other Python object into a stream of characters. The concept is that this character
stream provides all of the data required to recreate the object in another Python function.
The pickle data format is specific to Python This bears the advantage that the external
standards such as XDP and JSON impose no restrictions. Usually pickle uses a compact
binary representation. In our case we had to serialize our model which was in the form of a
jupyter notebook or ipynb file format. So we have to use the pickle.dump() function to load
or serialize the model in use. When deserializing the model or when we have to use the
model in our code again the pickle.load() function comes to use. Basically this function reads
the pickle representation of the object and returns the reconstituted object.

3.1.8 Heroku

It is a cloud based platform that has gained in popularity in recent years. Heroku is a popular
solution for many projects since it is so simple to utilise. It offers straightforward application
creation and deployment, with a special focus on facilitating customer-focused apps.

30
Businesses that use the Heroku platform may focus on refining their apps because the
platform takes care of the servers and platforms. Not to mention the infrastructure that keeps
them running. More time is dedicated to ensure that users have the best possible experiences.

3.2 Model Development

3.2.1 Dataset

The dataset was collected from Kaggle.com. Kaggle is a google owned subsidiary, web
community of information scientist and ML practitioners. Kaggle lets customers to find and
upload record units, discover and construct models via internet based record science
surroundings, engaging with other data scientists and also participate in competitions to
solve any data science problems. We are assured that the dataset is original and authentic
because kaggle is a trusted website that is used by over a million users worldwide.

Coming to the dataset we used. The name of the dataset is “Flight fare prediction dataset by
machine hack”. The dataset consists of a total entries of 10683. The total number of columns
or features in this are 11. Also when we take a closer look we find that out of the 11 variables
10 are categorical type of features and only 1 is of the numerical type. The numerical variable
is the price. As we have to predict the various prices so we will take the price as the
dependent variable and other 10 could be used as independent variables. Below mentioned
table shows the features along with their datatype and count.

Table 3.1 Dataset Information

# Feature Non Null Count Data type

0 Airline 10683 Object
1 Date_of_Journey 10683 Object
2 Source 10683 Object
3 Destination 10683 Object
4 Route 10682 Object
5 Dep_Time 10683 Object

31
 6 Arrival_Time 10683 Object
7 Duration 10683 Object
8 Total_Stops 10682 Object
9 Additional_Info 10683 Object
10 Price 10683 Int64

The “Airline” field contains all the names of the various flights. Some of these are also
repeated but come out as unique due to the different date of journeys etc. The “source” and
“destination” field as expected contains the starting point and the ending point for the
plane. The “dep_time” and the “arrival_time” contains the time stamp at which the plane
leaves the airport and lands at the desired airport. The “duration” field contains the time for
which the plane travels. The “total_stops” field includes the number of stops a flight takes
between the starting point and the ending point.

Also it is clear from the above stats that there are 2 Null values in the dataset. Those null
values are included in the column “Route” and the column “Total_Stops”. As it is our
responsibility to deal with those they were duly handled with.

Also during analysis we found that “Date_of_Journey” was of object data type so had to be
converted to timestamp so that it can be used correctly for prediction. Similarly for the ease
of use we extracted information from “Dep_Time” , “Arrival_Time”, and “Duration”
columns. Further doing the EDA we handled all the categorical data with the help of
encoders i.e OnehotEncoder for all the nominal data and LabelEncoder for the ordinal data.
The features that were handled here were “Airline”, “Source”, “Destination”, “Route”,
“Total_Stops”.

Below mentioned is the general information about the price feature of the dataset.

32
 Figure 3.16: Description of target variable

In this dataset the test set was given alongside the training dataset. It was a subset of the
same dataset. The test data set had a total of 2671 entries and 10 columns. The price was
columns was not there because as understood we have to predict that. The test set did not
have any NULL values. Again with the test set we performed all the work that we did with
the training data so that it was inline with the data we were training on.

3.2.2 Computational

For this project work we used the machine with the following specs at the time of training.

CPU: The computer we used had the following specs:

Table 3.2 CPU specifications

Parameter Specifications

CPU Model name Intel(R)Core(TM)

CPU frequency 2.3 Ghz

No of CPU cores 4

Available Ram 7.86 GB

Disk Space 400 GB

These are the CPU specs of the machine we used to do computations. Most of the
computational work mostly happens on the GPU. But CPU takes care of most of the
preprocessing. The large amount of RAM did not put loads of pressure and made it easier
for the whole dataset to be loaded in time and we did not have to worry about any system

33
crashes occurring. The clock speed of the CPU mentioned 2.3 Ghz is the baseic clock speed
which if needed can go upto 5 Ghz. But no overclocking was needed as the system was able
to do the work in its normal 4 cores.

Table 3.3: GPU Specifications

Parameter Specifications

GPU NVIDIA GeForce GTX 1060

GPU Memory 10 GB

GPU Memory Clock 1.40 Ghz

GPU Release Year 2016

Cores 2

Available RAM 6 GB

Disk Space 400 GB

Tools Used: We used the following tools in making our model.

 Python
 Matplotlib
 Seaborn
 Jupyter Notebook
 Pandas
 Plotly
 Numpy
 Scikit Learn
 Bootstrap
 HTML

34
  CSS
 Flask
 Heroku

These packages mentioned above were used in their latest upto date editions. The code
works properly and would not cause any issue until any further updates in them.

3.2.3 Experimental

Now we tried a few algorithms just to check them in comparison to what we used.

i. Linear regression
ii. Decision trees
iii. KNN

We are also providing with some basic definitions of the above algorithms so as to explain
easily and look why those algorithms did not perform the way desired.

i. Linear regression:
It is one of the well-known, understood and most used algorithm in ML and
statistical learning.
It is based on the concept of supervised learning. It performs the work to predict a
target variable (y) based on the independent variable (x), hence finds a linear
relationship between the two. The hypothesis function for linear regression is:
𝑦 = 𝜃1 + 𝜃2 𝑥

Where Θ1 is the intercept and Θ2 is the coefficient of x.

35
 Figure 3.16: Linear Regression

Once we find the best theta values we draw the best fit line and we would check for
our prediction. Our role is to minimize the cost function(J) which is given below
𝑛
1
𝐽 = 1 ∑(𝑝𝑟𝑒𝑑𝑖 − 𝑦𝑖 )2
𝑛
𝑖=1

𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 61%

ii. Decision trees:

This is an algorithm we already talked about . It uses a tree approach to get to the
results. In this algorithm we calculate entropy of every feature and the one with the
least is selected as the root node and we keep on splitting for the remaining columns.
The reason behind the failure of this algorithm is that he model works on data in a
sequential manner and the feature that has more domination outlasted the other
feature.

Figure 3.17: Decision Tree

𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 72%

iii. KNN:
KNN i.e K nearest neighbours is one of the algorithm that uses the information of its
neighbourhood to calculate results. This works on the concept that it checks
similarity between the new data and the available data and puts this new input into
the category which is most like to the other available categories. This algorithm
works for both regression and classification based problems. It is a non parametric
algorithm , which in simple terms mean that it does not make any assumptions on the

36
 given data. It does not learn from the data quickly but does the action on the dataset
at the time of need.

Figure 3.18: Depiction of KNN

𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 68%

3.2.4 Mathematical

Down below we have listed the formulas used by us in the report.

Linear regression: y = c + mx
𝑛
1
𝐽 = ∑(𝑝𝑟𝑒𝑑𝑗 − 𝑦𝑗 )2
𝑛
𝑗=1

Entropy: E(S) = -p(+) log p(+) – p(-)log p(-)

Information Gain: E(Y)- E(Y|X)

3.2.5 Statistical

As we have performed EDA on our dataset therefore we could infer some statistical insights
from that. We would be listing some of the graphs not all, that we visualized down below.

37
Also the below mentioned table contains some stats and information about the mean , min
etc of the numerical feature of our dataset.

Table 3.4 Dataset Stats

Price
Count 10683.0000
Mean 9087.064121
Std 4611.359167
Min 1759.0000
25% 5277.0000
50% 8372.0000
75% 12373.0000
max 79512.0000

As seen we can infer that the mean price for a flight ticket is around 9000 INR and the prices
can go as low as 1800 INR. So you can see the application of our model as it would be of
great use to an average customer as he/she can afford the ticket at a lower price than paying
higher amounts.

Figure 3.19: Journey_Month vs Price

We can see from the above graph that the fare price is fairly high in the month of March.
This could be due to the prior booking that people do for flights which means like for

38
travelling in may or june bookings can be done in march which leads to higher demands and
therefore higher prices.

Figure 3.20: Airline vs price

Here is another example where we can see that “jet airways” have the price at the highest
and other airways have a price almost at the same price bracket. These are a few examples
of the data visualization that can help in understanding the data far better.

39
Chapter 4 PERFORMANCE ANALYSIS

4.1 Analysis of System

Analysis of the system is the thing where we compare our system with the previous
implementations and check the performance of the system.In today’s world flight companies
try to change flight prices according to their needs and present condition. In this way they
can maximize their profits. Regression based solutions to this problem are common but they
come with their own disadvantage. With the technological advancements and the load
increase on these flight companies , regression does not give the desired results and offers
low accuracy. Customers on the other hand do not have a tool to have an idea about the ticket
prices that could help them in planning accordingly. Also many customers don’t know the
right time to book a ticket and sometimes have to pay more for a ticket. Consider an example,
the least expensive ticket will change its value over a period and its value may be high or
low depending on the time like summers ,winters etc or according to day ,night or evening.
This happens because the underlying goal of the plane carrier is to maximize its income. But
on the other side the purchaser try to get a ticket at a lower price so book them days before
the actual takeoff. But sometimes the price of a particular ticket does not fluctuate and he
consumer ends up paying more for the same. Also while working on the model the accuracy
score of our system on the training set was about 0.953 and on the test set was 0.805 which
indicates that our model was trained good and overfitting did not occur.

Given below are the results of different models used on this task before which includes the
models used by us.

Table 4.1: Models and precison

Model Name Accuracy

Linear Regression 0.619

Decision Tree 0.729

K-Nearest Neighbours 0.689

Random Forest (initial) 0.798

40
 Random Forest (after tuning) 0.813

Below are shown some of the output at various stage of this project.

Figure 4.1 : Original dataset

EDA

Figure 4.2: Departure time converted to single units

Figure 4.3: Arrival time converted to single units

Same was done for the duration field.

41
Data Preprocessing and visualization: Here we converted all the categorical data and did
some vizualizations.

Figure 4.4: Airline vs price

Figure 4.5: one hot encoding performed on the airline field

Figure 4.6: source vs price

42
Figure 4.7: Label encoding performed on the “total_stops” field

Figure 4.8: airline vs average price

Figure 4.9: Scatter plot of duration vs average price

43
 Figure 4.10: The original test set

All the actions that were performed on the training set were done on the test set.

Feature selection: To know which feature will have a better relationship with the target
variable

Figure 4.11: Graph of feature importance

44
Final output:

Figure 4.12: Scatter plot between the test and predicted values

Figure 4.13: Accuracy before hyperparameter tuning

Figure 4.14 Accuracy after tuning

45
 Figure 4.15 Flask App

Figure 4.16 Final Output

46
Chapter 5 CONCLUSIONS

5.1 Conclusions

Predicting air fare prices has been extensively studied and various methods and features have
been proposed for it to be performed. We gathered the flight price data from the web and
clearly showed that it is very much possible and feasible to predict the prices based on some
historical data. This report further shows that, ML predictors(models) are a more than
satisfactory option to know the air ticket prices. To the level of our understanding , maximum
of the preceding research work on the air plane price prediction concept focused on
traditional statistical procedures, which have their own obstacles of estimating and
prediction. Also we came to know that proper data and feature extraction and selection are
an important part of this process and helped us to come out with some helpful insights. Many
features were extracted form the data to make air travel segment easy to understand. We
from this project we can get information about the lows and highs of the ticket prices
depending on the time of day, the current day or even the weekends. We can also now with
full conviction draw a conclusion that if the model implemented in a proper manner can be
of great use in saving money of many people by providing them with the information about
the air ticket trends and given them an idea about the price which would help them to decide
whether to book a ticket right now or in the near future.

Also various ML models are studied and their efficiency and performance are compared so
as to get better results. This can also be attributed to the fact that the pricing models used by
different airlines are such that to maximize the revenue. Now with help of all this our model
can predict the values of the prices with and MSE.

47
5.2 Future Scope

The future scope can be extended to the fact that it can be used for predicting the prices of
the entire plane map of the airway service. More experiments on a larger air fare dataset is
an absolute requirement but this initial study depicts the potential of the ML models that can
be of use to both the user and the airline service. In the future our project can be prolonged
to encompass air ticket transaction stats, that can offer more detail on a selected itinerary ,
comprising of time and date of arrival/departure, seat area and much more. By aggregating
such records with the existing market section and the present day data , it is possible to build
an extra powerful and comprehensive price prediction system that can perform on a day by
day or even in hourly level. We can also use more better data ,trying with different features
to be able to get more efficient results. Additionally we can also look into various advanced
ML models, such as Deep learning models , even while running to enhance the present
models by hyperparameter tuning to attain a high quality structure.

48
 REFERENCES

[1] T.Wang et al., “A Framework for Airfare Price Prediction: A Machine Learning
Approach,” 2019 IEEE 20th International Conference on Information Reuse and
Integration for Data Science (IRI), 2019.

[2] K. Tziridis, T. Kalampokas, G.A Papakostas and K.I Diamantaras, “Airfare prices
prediction using machine learning techniques,” 2017 25th European signal
Processing Conference (EUSIPCO), 2017.

[3] S Rajankar, N Sakharkar, “A Survey on Flight Pricing Prediction using Machine

Learning”, International Journal of Engineering, 2019.

[4] Vinod Kimbhuane, Harshil Donga, Asutosh Trivedi, Sonam Mahajan and Viraj
Mahajan, “Flight Fare Prediction System”, May 2021

[5] Prashant Kapri, Shubham Patane, Noopur Thanvi, Rashmi Thakur, “Predictive
Model for Airlines Flight Delay and Pricing”

49