Car Resale Price Prediction App

Introduction

Dataset from Cardekho, leading car sale platform in India, contains information about used cars, with features like km_driven, fuel type, transmission type, number of previous owners and present retail price.

Objective

To predict the car resale price and to productionize the project to be used in real world.

Code and Resources

Python version: 3.7

Packages: pandas, numpy, matplotlib, seaborn, sklearn, pickle, flask

Kaggle source: https://www.kaggle.com/nehalbirla/vehicle-dataset-from-cardekho

Github repo: https://github.com/chekwei4/Car_Resale

Ipynb notebook: https://github.com/chekwei4/Car_Resale/blob/main/car_price_local.ipynb

Heroku App: https://car-price-app-1.herokuapp.com/

Model Building

Key steps:

Categorical variables were one hot encoded to give dummy variables.

Age of car feature was created by substracting current year with year which car was bought.

RandomizedSearchCV performed with RandomForestRegressor to give best hyperparameters.

Model Performance

R2: 0.92

MSE: 1.84

Distribution plot

y_test against y_pred

Scatter plot

y_test against y_pred

Front-End

Simple HTML file that allows user to input the key parameters required for model to predict.

Deployment

Key steps:

pickle the trained model
create new virtual environment and requirements.txt
wrap app in Flask API
deploy app onto Heroku, with github repo connection

Heroku App: https://car-price-app-1.herokuapp.com/

Credits

https://towardsdatascience.com/productionize-a-machine-learning-model-with-flask-and-heroku-8201260503d2

https://blog.usejournal.com/why-and-how-to-make-a-requirements-txt-f329c685181e

Youtube: Krish Naik, Ken Jee