https://github.com/animesh-algorithm/car-price-predictor-regression-analysis

Go through the README.md file.
https://github.com/animesh-algorithm/car-price-predictor-regression-analysis

Last synced: 3 months ago
JSON representation

Go through the README.md file.

• Host: GitHub
• URL: https://github.com/animesh-algorithm/car-price-predictor-regression-analysis
• Owner: animesh-algorithm
• Created: 2020-09-29T04:49:22.000Z (over 4 years ago)
• Default Branch: master
• Last Pushed: 2020-09-29T17:23:48.000Z (over 4 years ago)
• Last Synced: 2024-12-14T02:38:30.396Z (5 months ago)
• Language: Jupyter Notebook
• Homepage: https://car-price-predictor-1.herokuapp.com/
• Size: 350 KB
• Stars: 0
• Watchers: 2
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# Car Price Predictor App - Regression Analysis

## Table of Content

* [Demo](#demo)

* [Abstract](#abs)

* [Motivation](#motivation)

* [About the Data](#data)

* [Feature Engineering And Modeling](#model) 

  * [The Math Behind the Metrics Used](#metrics)

  * [Why I used Polynomial Features?](#poly)

  * [Why I converted the numerical data into log normal distribution?](#log)

  * [Why I perform One Hot Encoding on the data?](#ohe)

  * [Why I did feature scaling?](#fs)

  * [Training With Linear Models.](#linear_models)

  * [How I found the best model and their best hyperparameters using sklearn Pipelines and GridSearchCV?](#best_model)

* [Credits](#credits)

## Demo 

The final Output for the project can be viewed [here](https://car-price-predictor-1.herokuapp.com/)

![Demo](https://raw.githubusercontent.com/animesharma3/Car-Price-Predictor---Regression-Analysis/master/demo.png)

## Abstract 

* [Car Price Predictor](https://car-price-predictor-1.herokuapp.com/) is an **end-to-end Machine Learning Regression project**, built using **Flask** and deployed on **Heroku Cloud Platform**, which will help an **individual/dealer** to **predict the price** they can expect for any of their old car they would like to sell. The **dataset** on which the **ML models** are trained is taken from **kaggle**, provided by **cardekho.com**, has 4341 records and 8 dependent features.

* Various **ML models** are **tuned and trained** by making use of **sklearn pipelines and GridSearchCV**. Various **linear models, tree models and ensemble techniques** are employed for training (with all possible hyperparameters properly tuned). Out of them **Lasso(alpha=0.01)** happens to have better **r2-score** and least **Root Mean Squared Error.**

## Motivation 

My motivation for this project came from the **#66daysofdata** initiative started by **Ken Jee** from Youtube. If I haven't partipicated in this initiative, I may not be as accountable and consistent as I am in learning Data Science today.

## About the Data 

Our dataset has **4341 records** and **9 features**(including dependent feature).

### Brief Description about columns

 

  Columns

  Description

 

 

  Name

  Name of the cars

 

 

  Year

  Year of the car when it was bought

 

 

  Selling_Price

  Price at which the car is being sold

 

 

  Kms_Driven

  Number of Kilometres the car is driven

 

 

  Fuel_Type

  Fuel type of car (petrol / diesel / CNG / LPG / electric)

 

 

  Seller_Type

  Tells if a Seller is Individual or a Dealer

 

 

  Transmission

  Gear transmission of the car (Automatic/Manual)

 

 

  Owners

  Number of previous owners of the car.

 

 

  Name

  Name of the cars

 

## Feature Engineering and Modeling 

* ### **The Math behind the metrices used**

  * #### R2 Score

  

 

  * #### Mean Squared Error

  

* ### **Why I used Polynomial Features?**

  * Earlier for almost all the models, **r2_score** is close to **0.73**

  * But after increasing the no. of independent features to 45 using **sklearn.preprocessing.PolynomialFeatures** of **degree 2** the **r2_score** increases to **0.94** i.e

**Our Model is fitting better**, which is evident from the graphs below, which is basically plotted to show the **variance** between the actual and predicted value of Selling Price




   

         Before Applying Polynomial Features            |            After Applying Polynomial Features

       

  

 * ### **Why I converted my data into log normal distribution?**

   * One of the basic assumption when you are training a **Linear Regression** model or compiling an **Artificial Neural Network** is that your data is **normally distributed**. I plotted the graph, I got some bell curve but I was not sure whether it is perfectly normally distributed. So I performed the **log normal distribution** for all my numerical data.

          Before Applying Log Normal Dist.               |            After Applying Log Normal Dist.

       

 * ### **Why I perform One Hot Encoding on the data?**

   * My Data have some categorical features like **Fuel Type, Seller Type and Transmission**, which doesn't make sense to any Machine Learning or Deep Learning Models. Hence they needed to be converted into numerical data and one way to do that is through **One Hot Encoding**, which I did using **pd.get_dummies** function.

 

 * ### **Why I did feature scaling?**

   * All of my features have different units, ex. kilometers, no. of years, rupees etc. So I scaled all of my features.

   * I tried both **MinMaxScaler()** and **StandardScaler()**, the latter one worked better.

 * ### **Training with linear models**

   * I trained my dataset with various linear models for ex. **Linear Regression, Lasso Regression, Ridge Regression and Bayesian Ridge Regression**.

   * To my surprise all models performed same way with nearly same value of **r2_score, Mean Absolute Error and Mean Squared Error**.

   

                      Linear Regression               |            Lasso Regression

           

   

                      Ridge Regression                |             Bayesian Ridge Regression

        

 * ### **How I found the best model and their best hyperparameters using sklearn Pipelines and GridSearchCV?**

   * I made use of **sklearn pipelines** to its best. I created multiple pipelines with respect to the model and their hyperparameters. 

   * Then I combined the power of **sklearn pipelines** and **GridSearchCV**, which trained all the models for me, with all the possible hyperparamerter and finally gave the **best model** which in our case happens to be **Lasso(alpha=0.01)**.

   * I **dumped** that model using **pickle** library and then **loaded** the model into the **Flask** app.

   * The best model gave the following scores - 

       * R2 Score  - 0.9442452908334493

       * Mean Absolute Error -  0.7422001552130918

       * Mean Squared Error -  1.2843423154102787

   * And the following graph for the **variance in Predicted and Actual Value of Selling Price**

    

        

* For more deep dive into the modelling stuff you can check out the [Ipython Notebook for Feature Engineering and Choosing Best Model](https://github.com/animesharma3/Car-Price-Predictor---Regression-Analysis/blob/master/Feature%20Engineering%20and%20Choosing%20Best%20Model.ipynb)

## Credits 

Special thanks to 

* **cardekho.com** for providing the dataset

* **Google Colaboratory** for providing the computational power to build and train the models.

* **Heroku** for making the deployment of this project possible.

* **#66daysofdata** community created by **KenJee** that is keeping me accountable in my Data Science Journey.