CITC_Task

Cars price prediction

This repo is for a job offer coding task and the task was as follow:

You need to build a model that predicts the price of the car using Python. Please make sure to use at least 3 different methods/algorithms to predict the price and report which one will be most accurate.

Dataset:

the dataset contains 8 features 301 row

1. Car Name: Nominal

Missing	Unique
1	47

2. Year: Numeric

Min	Max	Mean	StdDev
2003	2018	2013	2.891

3. Mileage: Numeric

Min	Max	Mean	StdDev
500	500000	36947.206	38886.884

4. Fuel type: Nominal

Label	Count
Petrol	239
Diesel	60
CNG	2

5. Seller Type: Nominal

Label	Count
Dealer	195
Individual	106

6. Transmission type: Nominal

Label	Count
Manual	261
Automatic	40

7. Previous owners: Numeric

Min	Max	Mean	StdDev
0	1	0.043	0.248

8. Selling price: Numeric

Min	Max	Mean	StdDev
0.1	35	4.643	5.081

Bellow a histogram for all features

Preprocessing

• Remove Outliers
• Fill the Missing value
• Convert to numeric representation
• Add new features
• Handel Car's name
• Feature selection

Correlation matrix

Models

• Shallow learning (Machine Learning)
  • Linear regression
  • Random forest regressor
  • Support vector regression
• Deep learning
  • CNN
    • Baseline model
    • Standardizing Dataset
    • Tuning the network

Conclusion

Preprocessing	Models	MSE	R2	RMSE	Mean
Without Names	LR	4.096	0.668	4.096
One-Hot-Encoding	LR	4.105	0.731	4.105
Without Names	RFR	4.205	0.660	4.205
One-Hot-Encoding	RFR	2.723	0.821	2.723
Without Names	SVM	12.037	-0.026	0.026
One-Hot-Encoding	SVM	15.883	-0.042	15.883
One-Hot-Encoding	CNN	22.26			-27.59
One-Hot-Encoding	CNN	4.91			-4.05
One-Hot-Encoding	CNN	5.26			-4.19

As we can see the Random Forest Regression gives a high R2 score 82% with Mean square error 2.723 that is mean it is the closest to the line best fit.

Visualizations Excerpts

The plot of the Random forest regressor model which achieved the highest accuracy

Graph shows predictions VS the actual which miss the actual values at some places

R-Squared & RMSE by number of features

The jupyter notebook has the following sections:

• Libraries
• Load dataset
• Exploratory Data Analysis
  • Selling price
  • Convert to numeric representation
  • Missing values
    • year
    • Mileage
    • Care Name
  • Feature Extraction
    • Correlation matrix
• Split the data (training\testing)
• Models
  • Shallow Learning (ML)
    • Linear regression
      • Feature Selection using RFE & K-Fold Cross Validation
    • Random force Regressor
      • Feature Selection using RFE & K-Fold Cross Validation
    • Support vector regression
  • Deep learning
    • CNN
• Conda setup