Flight Price Prediction Using Machine Learning

Ankita Panigrahi1, Rakesh Sharma2, Sujata Chakravarty3, Bijay K. Paikaray4 and

Harshvardhan Bhoyar 5

123
Dept. of CSE, Centurion University of Technology and Management, Odisha, India.
4
School of Information & Communication Technology, Medhavi Skills University, Sikkim, India
5
Faculty. of Management Studies, Sri Sri University, Odisha, India

Abstract
Currently, everyone loves to travel by flights. Going along with the study, the charge of
travelling through a plane change now and then which also includes the day and night time.
Additionally, it changes with special times of the year or celebration seasons. There are a few
unique elements upon which the cost of air transport depends. The salesperson has data
regarding each of the variables, however, buyers can get confined information which is not
sufficient to foresee the airfare costs. Considering the provisions, for example, time of the
day, the number of days remaining and the time of take-off this will provide the perfect time
to purchase the plane ticket. The motivation behind this paper is to concentrate on every
component that impacts the variations in the costs of this means of transport and how these
are connected with the diversity in the airfare. Subsequently, at that point, utilizing this data,
construct a framework that can help purchasers when to purchase a ticket. Machine Learning
algorithms prove to be the best solution for the above-discussed problems. In this project,
there is an implementation of Artificial Neural Network (ANN), LR (Linear Regression), DT
(Decision Tree), and RF (Random Forest).

Keywords
Machine Learning Algorithms, airfare, supervised learning, predictions, flight, Linear
Regression, Artificial Neural Network, Random Forest.

1. Introduction
A person who already has reserved a ticket for a flight realizes how powerfully the price of the ticket
switches [1]. Airline utilizes progressed techniques considered Revenue Management to accomplish a
characteristic esteeming technique [2]. The most affordable ticket available changes over a course of time.
The expense of the booking may be far and wide. This esteeming technique normally alters the cost
according to the different times in a day namely forenoon, evening, or night. Expenses for the flight may
similarly alter according to the different seasons in a year like summers, rainy and winters, also during the
period of festivals. The buyers would be looking for the cheapest ticket while the outrageous objective of
the transporter would be generating more and more revenue. Travelers for the most part attempt to buy
the ticket ahead of their departure day. The reason would be their belief that the prices might be the highest
when they would make a booking much nearer to the day of their flight but conventionally this isn't
verifiable. The buyer might wrap up paying more than they should for a comparable seat. Considering the
challenges faced by the travellers for getting an affordable seat, various strategies are utilized which will
extract a particular day on which the fare will be the least. For this purpose, Machine Learning comes into
the picture. Gini and Groves developed a model using PLSR, to predict the appropriate time to book the
seats [3]. They extracted their data from well-known booking websites from 22/02/2011 to 23/06/2011.

ACI’22: Workshop on Advances in Computation Intelligence, its Concepts & Applications at ISIC 2022, May 17-19, Savannah, United States
EMAIL: [email protected] (A. 1); [email protected] (A.2); [email protected] (A.3); [email protected]
(A. 4); [email protected] (A. 5)
ORCID: 0000-0001-5843-0335 (A. 4)
©️ 2020 Copyright for this paper by its authors.
Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR Workshop Proceedings (CEUR-WS.org)

172
Using the Linear Quantile Blended Regression methodology, Janssen [4] developed an assumption model
for the route of San Francisco to New York with already available data on flight fares for each day provided
by www.infare.com. The two important features were the day count from departure and which day of the
week it is, whether it’s weekday or weekend. This model was capable enough to predict the expense for
the flight for the days that were nowhere close to the day of departure but the results were not satisfying
if it would be close to the date of journey. A ticket-purchasing time incremental model depending upon
marked point processors and information extracting systems and computable investigation strategy was
suggested by Wohlfarth [5]. The proposed system changes the heterogeneous value arrangement
information to added value arrangement system. For choosing the best synchronizing group and later
comparison of advancement model a tree-based order calculation has been used. Papadakis [6] anticipated
whether there would be a fall in the airfare later on by addressing the issue as a classification task using
Logistic Regression, Linear SVM and Ripple Down Rule Learner models. Ren, Yang, and Yuan [7]
worked on Linear Regression, Naïve Bayes, SoftMax Regression, and SVM models in predicting the
prices.

2. Data Collection
The assortment of data is the very first step in machine learning projects. There are various sources of
data available on numerous websites that are deployed to construct the models. These sites supply a huge
variety of data regarding different airlines, routes, times, and tolls. In this part, data gathered from the
various available sources are studied. For the execution of this, information is brought from a site called
Kaggle. For the assortment of the data and to execute the model's Python is utilized [8-15]. The dataset
collected contains information about different airlines in India. It consists of various factors which affect
the price of a flight ticket including the price for a particular flight. It contains 10683 rows of data. The
features present in the dataset are the name of companies, Date of travelling, Origin, terminus, path of
travelling, Time of Departure, Time of Arrival, Travelling Hours, Total Stoppage, Additional Info, and
Price.

3. Cleaning and Preparing of Data

Cleaning and preparing data are a very important step in machine learning. The data collected can’t
be used raw as it may contain certain parameters which would be of no use and also certain data can’t be
used the way it would be present in the dataset. So, before proceeding to the actual work, the data needs
to be filtered and it should be absolutely clean. For achieving this, all the duplicate and null values are
removed from the dataset and specific data is converted to a usable format.

4. Machine Learning Techniques

Various conventional machine learning algorithms are used for creating a model for flight fare
prediction which is ANN, LR, DT, and RF. These loads of machine learning techniques are executed using
the sci-kit-learn library available in python. For assessing the exhibition of these algorithms, definite
boundaries are thought of. These are mentioned as follows: MAPE (Mean Absolute Percentage Error) and
RMSE (Root Mean Square Error).

4.1 RMSE
RMSE is a tool that helps in determining how accurately the model is making the predictions. It
calculates how much error the model creates while making these predictions. It measures the standard of
predictions. Mathematically, it is defined as the square root of the average of the squares of all the errors.
Error is defined as the difference between the actual and predicted value. Less the RMSE, the better the
performance of the model is. Usually, an RMSE score of less than 1 is considered the best.

173
 (1)

4.2 MAPE
Mean Absolute Percentage Error is most often used in regression problems. It is most popular in
calculating errors in forecasting. It gives an idea about how much accurately the model is evaluating the
predictions. Statistically, it is the mean or average of the absolute percentage errors of forecasts. Error is
characterized as the contrast between the actual and predicted value. Less the MAPE, the better the
exhibition of the model is. Typically, a MAPE score of below 1 is viewed as awesome.
𝒏
𝟏𝟎𝟎 𝑨𝒕 − 𝑭 𝒕
𝑴𝑨𝑷𝑬 = ∑| |
𝑵 𝑨𝒕
𝒕=𝟏 (2)
Here,
At is the actual-value
Ft is the forecasted-value
5. Machine Learning Algorithms Used
5.1 Artificial Neural Network (ANN)
An artificial Neural Network is simply a Neural Network that resembles the biological Neural
Network present in the human brain. It is designed in a way such that it would function the same way a
human brain function. It is the collection of millions and millions of artificial neurons. These artificial
neurons are the building blocks of the ANN model. Artificial Neuron consists of Inputs and their
corresponding weights. An activation function is chosen which takes these inputs multiplies them to their
corresponding weights and produces the output. Every Artificial Neural Network must have three layers:
the input layer which takes the input, the hidden layer where all the computations take place, and the
output layers which produce the output.

Figure 1. Flow Diagram of ANN Model

In the case of hidden layers, we have used Relu as Activation function with 20 and 10 for weights
whereas Linear Activation Function with weight 1 is used in case of the final output. Here, adam optimizer
is used.
𝑁𝑗−1
(3)
𝑗−1
𝑍𝑖 = ( ∑ 𝑋𝑘 𝑊𝑘,𝑖 − 𝑏𝑘 )
𝑘=1

174
 1
𝑓(𝑍𝑖 ) = (4)
1 + 𝑒 −𝑍𝑖

5.2 Linear Regression

Linear Regression is an algorithm in machine learning. It works by finding the relationship between single
or multiple input variables and the output variable. These relationships are built with linear predictor
functions. The graph of a linear regression model is linear justifying its name.
𝑦(𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑) = 𝑏0 + 𝑏1 ∗ 𝑥 (5)
Here,
y is dependent variable,
x is independent variable,
b0 is constant,
b1 is slope.

5.3 Decision Tree

This model is a member of a supervised learning family. It can fit well in both classification and regression
problems. As its name says, it is structured like a tree containing the decision nodes and leaf nodes.
Decision nodes have multiple branches for decision making where leaf nodes represent the outcomes of
these decisions which is further not divided into any branches.

ROOT NODE Branch/Sub-Tree

Splitting

A Decision Node
Decision Node
B C

Terminal Node Terminal Node

Terminal Node Decision Node

Terminal Node Terminal Node

Figure 2. Decision Tree Process

If we write mathematically,
Entropy having 1 attribute:
𝑐

𝐸(𝑆) = ∑ −𝑝𝑖 𝑙𝑜𝑔2 𝑝𝑖 (6)

𝑖=1
Entropy having multiple attributes:
𝐸(𝑇, 𝑋) = ∑ 𝑃(𝑐)𝐸(𝑐)
(7)
𝑐∈𝑋

175
5.4 Random Forest
Like the Decision Tree, RF is also a supervised learning technique. Random forest works with multiple
decision trees. Here, the trees are operated as an ensemble. Every tree present in a random forest divides
a class prediction and the class having the most votes comes out as models’ prediction.

6. Algorithms Evaluation
On comparing the Root Mean Square Errors of the pre-processed data when applied on proposed
algorithms it is specified that Artificial Neural Network gives 0.008410 followed by Random Forest
giving 0.006240 then Linear regression with 0.006109 closely baking up by Decision Tree with the least
error of 0.006101 which shows the Decision Tree model works more precisely than others when applied
on the given data. The value given in Table 1 is graphically represented in Figure3.
Table 1
Different ML Models RMSE Errors
ML Algorithms RMSE
Artificial Neural Network 0.008410309713082834
Linear Regression 0.006109087698177261
Decision Tree 0.0061019746207730645
Random Forest 0.0062402313794453

RMSE RESULTS
0.01
RMSE Values

0.008
0.006
0.004
0.002
0
Artificial Neural Linear Regression Decision Tree Random Forest
Network
Algorithm Applied

Figure 3. Result Analysis of RMSE for all applied Models

If we go through the Mean Absolute Percentage Error the results show that again the Decision Tree
got the least Error when compared with all other models. The value given in Table 2 is graphically
represented in Figure 4.
Table 2
Different ML Models MAPE Errors
ML Algorithms MAPE
Artificial Neural Network 0.6831663296497983
Linear Regression 0.5202171579180117
Decision Tree 0.5202012748283965
Random Forest 0.5291447939343533

176
 MAPE RESULTS

0.8
MAPE Values 0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
Artificial Neural Linear Regression Decision Tree Random Forest
Network

Algorithm Applied

Figure 4. Result Analysis of MAPE for all applied Models

7. Conclusion
We learn that ML models can be used to predict prices based on earlier data more correctly. The
presented paper reflects the dynamic change in the cost of flight tickets from which we get the
information about the increase or decrease in the price as per the days, weekends, and the time of the
day. With the Ml algorithm applied on various datasets, better results can be obtained for prediction.
The error values that we got for Artificial Neural Network are comparatively high but for obtaining
lesser values we can use evolutionary algorithms of ANN like genetic algorithms in the future.

8. References
[1] Rajankar, Supriya, and Neha Sakharkar. "A Survey on Flight Pricing Prediction using Machine
Learning." Internatıonal Journal Of Engıneerıng Research & Technology (Ijert) 8.6 (2019): 1281-
1284.
[2] Smith, Barry C., John F. Leimkuhler, and Ross M. Darrow. "Yield management at American
airlines." interfaces 22.1 (1992): 8-31.
[3] Groves, William, and Maria Gini. "An agent for optimizing airline ticket purchasing." Proceedings
of the 2013 international conference on Autonomous agents and multi-agent systems. 2013.
[4] Janssen, Tim, et al. "A linear quantile mixed regression model for prediction of airline ticket prices."
Radboud University (2014).
[5] Wohlfarth, Till, et al. "A data-mining approach to travel price forecasting." 2011 10th International
Conference on Machine Learning and Applications and Workshops. Vol. 1. IEEE, 2011.
[6] Papadakis, Manolis. "Predicting Airfare Prices." (2014).
[7] Ren, Ruixuan, Yunzhe Yang, and Shenli Yuan. "Prediction of airline ticket price." University of
Stanford (2014).
[8] Tziridis, Konstantinos, et al. "Airfare prices prediction using machine learning techniques." 2017
25th European Signal Processing Conference (EUSIPCO). IEEE, 2017.
[9] Boruah, Abhijit, et al. "A Bayesian Approach for Flight Fare Prediction Based on Kalman Filter."
Progress in Advanced Computing and Intelligent Engineering. Springer, Singapore, 2019. 191-203.
[10] S. Chakravarty, B. K. Paikaray, R. Mishra and S. Dash, "Hyperspectral Image Classification using
Spectral Angle Mapper," 2021 IEEE International Women in Engineering (WIE) Conference on

177
 Electrical and Computer Engineering (WIECON-ECE), 2021, pp. 87-90, doi: 10.1109/WIECON-
ECE54711.2021.9829585.
[11] Wang, Tianyi, 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). IEEE, 2019.
[12] Abdella, Juhar Ahmed, et al. "Airline ticket price and demand prediction: A survey." Journal of
King Saud University-Computer and Information Sciences 33.4 (2021): 375-391.
[13] Zhao-Jun, Gu, Wang Shuang, and Zhao Yi. "Flight ticket fare prediction model based on time-
serial." Journal of Civil Aviation University of China 31.2 (2013): 80.
[14] Huang, Tenghui, Chih-Chien Chen, and Zvi Schwartz. "Do I book at exactly the right time?
Airfare forecast accuracy across three price-prediction platforms." Journal of Revenue and Pricing
Management 18.4 (2019): 281-290.
[15] S. Chakravarty, P. Mohapatra, P. K. Dash, (2016), Evolutionary Extreme Learning Machine for
Energy Price Forecasting, International Journal of Knowledge-Based and Intelligent Engineering
Systems, 20, 75-96
[16] https://www.kaggle.com/nikhilmittal/flight-fare-prediction-mh/
[17] https://github.com/rishabdhar12/Flight-Price-Prediction/tree/main/Dataset

178