Used Cars Price Prediction

DSCI-522 - Group 308 Authors: Andrés Pitta, Braden Tam, Serhiy Pokrovskyy

About

In this project we attempt to build a regression model which can predict the price of used cars based on numerous features of the car. We tested the following models: support vector machine regression, stochastic gradient descent regression, linear regression, K-nearest neighbour regression as well as random forest regression and gradient boosted trees. We found that support vector machine regression shown the best results, producing an score of 0.877 on the training set, score of 0.832 on the validation set and score of 0.830 on the test set. The training and validation scores are computed from a very small subset of the data while the test score used a much larger subset. Given that the dataset was imbalanced by manufacturers, this led to a bit worse prediction of the classes that were quite sparse because the model was not able to learn enough about those classes in order to give good predictions on unseen data.

The data set used in this project is "Used Cars Dataset" created by Austin Reese. It was collected from Kaggle.com (Reese 2020) and can be found here. This data consists of used car listings in the US scraped from Craigslist that contains information such as listed price, manufacturer, model, listed condition, fuel type, odometer, type of car, and which state it’s being sold in.

Report

The final report can be found here.

Usage

To replicate the analysis, clone this GitHub repository and follow these instructions:

1. Reset the repository and clean any prebuilt results:

make clean

2. Run the complete pipeline:

make all

Quick Analysis

The original dataset’s size is 1.35GB. Fitting the model may take many hours (seriously!) Thus, we provided a convenient method to replicate a quick version of the model with just a portion of the data:

1. Reset the repository (without removing the data file data/vehicles.csv):

make partial_clean

2. Run the quick version of the pipeline:

make quick TRAIN_SIZE=0.01

You may choose other percentage value (0-to-1) For 1% (TRAIN_SIZE=0.01) expected runtime is 5 minutes. Keep in mind that lower dataset size decreases accuracy. Also, train / test metrics are automatically embedded in generated reports, thus the reports will reflect those metrics and not the best ones we had prebuilt by default.

Using Docker

If you do not want to install all the software dependencies, you may want to use Docker. Running the Docker pipelines will download a prebuilt Docker container image for our project. Be advised, that the size of container is approximately 1.86 GB, plus the original data file of 1.35 GB will also be downloaded afterwards.

To replicate the quick version of this analysis using Docker run the following commands (on Windows use git bash):

1. Reset the repository (without removing the data file data/vehicles.csv):

make partial_clean

2. Run the quick version of the pipeline using Docker:

make quick_docker

Similarly, you may use make all_docker to build full model with Docker (same advisory on time / resources applies)

Makefile dependency maps

Please consider the following dependency maps for the make processes:

1. The dependencies for the complete version of this pipeline are shown in the following diagram:

2. The dependencies for the quick version of this pipeline are shown in the following diagram:

Dependencies

NOTE When installing the following packages, conda install is preferred:

• Python 3.7.3 and Python packages:
  • altair==3.2.0
  • selenium==3.141.0
  • docopt==0.6.2
  • pandas==0.24.2
  • numpy==1.16.4
  • statsmodel==0.10.0
  • plotly==4.3.0
  • plotly-orca==1.2.1
  • scikit-learn==0.20.4
• R version 3.6.1 and R packages:
  • knitr==1.24
  • docopt==0.6.1
  • tidyverse==1.3.0
  • readr==1.3.1

Notes

Please be advised that in order to reproduce the analysis, the scripts/download.py script will download approximately 1.35GB original data file. Also, training the full model with scripts/train_model.py may take several hours. You may consider using command line arguments to train on a configurable subset of training data (which may affect the trained model) Please consult usage data for each script for further details and command line arguments.

License

The Used Cars Prediction materials here are licensed under the Creative Commons Attribution 2.5 Canada License (CC BY 2.5 CA). If re-using/re-mixing please provide attribution and link to this webpage.

References