Flight Price Predictor: Project Overview

• Built a model to predict flight prices depending on various user inputs and deployed it on flask
• Trained 10 different models to get the best performing model and optimized it for even better results
• R2 score of the final trained model is 80.97%

Table of Contents

• Code and Resources Used
• Directory Tree
• Data Preprocessing
• EDA
• Feature Engineering
• Data Cleaning
• Feature Selection
• Model Building
• Hyper-parameter Tuning
• Deployment

Code and Resources Used

To install the required packages and libraries for this project, run this command in the project directory after cloning the repository:

pip install -r requirements.txt

Dataset: Download the entire dataset from

Directory Tree

├── Dataset and Images 
│   ├── Data_Train.xlsx
│   ├── model.png
│   ├── plane.jpeg
├── Templates
│   ├── final_trained_model.pkl
│   ├── home.html
│   ├── plane.jpeg
├── Flight Price Prediction.ipynb
├── README.md
├── app.py
├── final_trained_model.pkl
├── requirements.txt

Data Preprocessing

Following changes were made to the data to make it usable for a model:

• Column with Null Values was removed.
• Data values for 'Delhi' and 'New Delhi' were combined.
• Date and Duration which was present as string values was converted into timestamp format.

EDA

Following analysis were made related to dataset:

• What time of the day most flights take off
• If the duration of flights affect its price
• If total number of stops affect the flight price
• Ticket Fare Distribution by Airline
• Median ticket fare by Airline

Feature Engineering

• One-Hot Encoding and Label Encoding was done to convert categorical dataset into vector
• Target Guided Encoding was done to avoid curse of dimensionality during feature encoding
• A one hot encoder library was used whereas a label encoder was made manually

Data Cleaning

• The unwanted columns for the model were removed
• Outlier range and the outliers were detected using IQR method
• The outliers were replaced with the median of the remaining data values

Feature Selection

• Mutual Information Regression was used to identify dependency between the variables to select the best features for the model
• As all the features showed a good dependency with the target variable no specific feature was selected

Model Building

The data was split into 75% training and 25 % test set. An automated ML model was made so that mutiple models can be evaluated in a single code

10 different models were tried and evaluated based on their metrics:

• Random Forest Regression : R2 score = 79.69%
• Decision Tree Regressor : R2 score = 64.87%
• Linear Regression : R2 score = 60.77%
• Ridge Method : R2 score = 60.77%
• Lasso Method : R2 score = 60.77%
• ElasticNet Method : R2 score = 57.27%
• Support Regression : R2 score = 2.62%
• K-NN : R2 score = 64.77%
• MLP Regressor : R2 score = 56.71%
• Huber Regressor : R2 score = 59.4%

Hyper-parameter Tuning

Clearly Random Forest outperforms the other methods but its performance can be still improved. RandomizedSearchCV was used to find the hyper-parameters and optimize the model upto an R2 score of 80.97%

Deployment

A Final Trained model was built on Random Forest regression and deployed on flask as a web app. The Final model can be downloaded from final_trained_model.pkl