In this project, we use several ML methods (linear regression, ridge regression, random forest, and neural networks) to build models, predicting second-hand car price, based on "used_car_train_20200313.csv" dataset.

This project has the following parts.

1. Data cleaning & Feature engineering
2. Build Machine Learning models (linear regression, ridge regression, random forest)
3. Build Deep Learning models (neural networks)
4. Final analysis and report (Note: the complete project description, process, results analysis, and charts are in report pdf)

codes

• "data_process.py" do all the Data cleaning & Feature engineering work.
• "machine_learning.py" do all the model training work.
• "main.py" operate all the functions in "data_process.py" and "machine_learning.py".
• "more_test.py" do further comparison among trained models.

output (plot & results) Store all generated plots, trained models parameters, and validation results.

part2 (final) Report Store complete project desciption, process, results analysis, and charts.

Dataset we use: used_car_train_20200313.csv

Details about this dataset:

1. 150000 rows/samples, 31 columns/attributes.
2. Attributes in our dataset: ['SaleID', 'name', 'regDate', 'model', 'brand', 'bodyType', 'fuelType', 'gearbox', 'power', 'kilometer', 'notRepairedDamage', 'regionCode', 'seller', 'offerType', 'creatDate', 'price', 'v_0', 'v_1', 'v_2', 'v_3', 'v_4', 'v_5', 'v_6', 'v_7', 'v_8', 'v_9', 'v_10', 'v_11', 'v_12', 'v_13', 'v_14']
   Among them, v_0 to v_14 are anonymous variables .
3. The whole dataset has information about 150000 used cars, I will use 100000 for training set, 50000 for test set.

1. Use normal polynomial interpolation to fill missing values.
2. Use heatmap to identify top 10 attributes most correlated/relevant to the current price
3. Use pearson correlation coefficient to compare each individual attribute/column with the price attribute/column, check whether strong correlation exists.
4. Choose 3 attributes/factors from those 10 factors, using trendline to draw graphs of each attribute vs. price, observing the correlation.

1. Train a linear regression model.
2. Train a ridge regression model.
3. Train a random forest model.
4. Store mean squared error & learning curve, compare mean squared error of each model, choose the best model.

1. Train a neural net with two hidden layers, sigmoid as activation function, L2 normalization in the loss function, and learning rate between 0.1-0.2.
2. Train a neural net with two hidden layers, relu as activation function, L1 normalization in the loss function, and learning rate between 0.1-0.2.
3. Train a neural net with two hidden layers, relu as activation function, L2 normalization in the loss function, and learning rate between 0.1-0.2.
4. Store mean squared error & learning curve, compare mean squared error of each model, choose the best model.

To all the analysis and training described above, their results, charts, and analysis are in report pdf.

Please follow steps listed below one by one, from top to bottom. Please put all python files, including the dataset, into the same folder!!! (for convenience of description, we call this folderX)

Dowload dataset from this link below, and put it in folderX. https://drive.google.com/file/d/1fgUXVEAmnvWJVNXFKqOrgHTLkgG66Hoj/view?usp=sharing

Put all my submitted python files ('data_process.py', 'machine_learning.py', 'main.py', 'more_test.py') in folderX.

Open your terminal, use pip to install these libraries:

1. pandas
2. scikit-learn
3. numpy
4. matplotlib
5. seaborn
6. tensorflow

VScode terminal path setting: When operate the code, please use the "Python" terminal, and set the path to the direct folder where you put / store these code files (folderX). See the Red Highlighted part in example given in page 18, part 10a) of my report.

1. Open and Run main.py
2. Open and Run more_test.py
3. Note: make sure all files are in the same folder and the path is set correctly.

For meaning that each output file represents, please see page 18-19, part 10b) in my report.