Authors: Lisa Sequeira, Alex Taciuk, Peng Zhang, Gaoxiang Wang
Our research question: Given attributes of car models from the 90s, can we predict how popular each car model will be?
In order to pay for MDS tuition, our team is starting a side-business buying and selling used cars. Used cars have seen recent high demand and varied prices so we think this is a fertile ground for a price arbitrage business. In order to run a successful business buying and selling used cars, we need to determine the most popular used car models.
The cars we are aiming to buy and sell will be older (ideally from the 90s), so that we can afford to buy and sell them. We will use a dataset of car popularity by model primarily collected in June, 1997 by Marko Bohanec to evaluate HINT (Hierarchy INduction Tool), and is presented in B. Zupan, M. Bohanec, I. Bratko, J. Demsar: Machine learning by function decomposition. ICML-97, Nashville, TN. 1997. The data can be accessed from the UC Irvine Machine Learning Repository and follows a Creative Commons Attribution 4.0 International License. Each row of this dataset is a car model, with popularity.
Using the above dataset, we will create a model to predict how popular cars were among drivers, that is generalizable to older car models beyond this dataset. Specifically, we want to predict the correct category (or classification) of car popularity/acceptability (class) based on six commonly available car attributes (features):
- Price
- Price of maintenance
- Number of doors
- Capacity of the car (number of passengers we can carry)
- Size of trunk or luggage space
- Estimated safety of the car
The first thing we did is split our data into test and training sets (with a 90% to 10% split). We then performed exploratory data analysis (EDA). We reviewed the distribution of our target class (to see if we have a class imbalance) and compare each of the features to the target class to see if there are any relationships or strong correlations between them. We also evaluated how “dirty” our data is so that we know what types of transformations we needed to perform. Using our initial EDA also helped us make an assessment on the models we wanted to use.
Next we began pre-processing and data cleaning on our training data by creating a column transformer to add into our pipeline. Given our features contain a lot of categorical variables and the final predicted target is categorical, we used and compared the performance of decision tree classifier, RandomForest and Multinomial NB models from SciKit Learn. We used the DummyClassifier as our baseline model. For each of our selected models we conducted hyperparameter optimization and compared each model using a balanced accuracy score to select the optimal model for downstream testing. This was presented as a table and graphs in the analysis.
After selecting our optimal model (RandomForest Classifier), we used the pipeline created above to re-run these on our test dataset and analyzed the scores from the prediction to decide if our model is good enough to use on deployment data.
To start off, fork the repo and clone the forked repo to your local machine by running:
git clone https://github.com/UBC-MDS/Car-Acceptability-Prediction.git
First, you will need to install Docker.
Next, clone the GitHUb repository.
Finally, via navitgate to the root of the repository and run the following command:
docker run --rm -v /$(pwd):/home ataciuk/dockerfile-practice:84645833 make -C /home all
To 'clean' the repository, or remove the produced files, run this command:
docker run --rm -v /$(pwd):/home ataciuk/dockerfile-practice:84645833 make -C /home clean
Note: If you are using Apple silicon, you will need to add --platform linux/amd64 after the --rm.
The Docker image is hosted in this repository.
Run the code block below to install the environment and activate it, before proceeding with the steps below.
conda env create -f env522car.yaml
conda activate 522env
To reproduce the analysis, run the command at the root directory.
make all
To reset and redo the analysis from beginning, run the command at the root directory.
make clean
Run the code block below to install the environment and activate it, before proceeding with the steps below.
conda env create -f env522car.yaml
conda activate 522env
- Using
docoptin Python to download data file from external URL
To proceed, run the following command at the command line and set root directory as the current working directory.
python src/download_data.py --url=https://archive.ics.uci.edu/ml/machine-learning-databases/car/car.data --local_file_path=data/raw/raw_data.csv
- Using
Altairin Python to render the literate Exploratory Data Analysis.
To proceed, open and run src/car_popularity_analysis.ipynb
- For processing raw data, run the following command at the root directory.
python src/data_processing.py --raw_data_path=data/raw/raw_data.csv --clean_data_folder_path=data/processed
- For eda plotting of training data, run the following command at the root directory.
python src/eda_car_popularity.py --training_data_path=data/processed/training.csv --folder_result_path=result
- For machine learning analysis for traning data, run the following command at the root directory.
python src/car_classifier_analysis.py --train_test_folder_path=data/processed --result_folder_path=result
- To render final report, run the following command at the project directory.
Rscript -e "rmarkdown::render('doc/car_popularity_prediction_report.Rmd')"
The conda environment, called env522car.yaml, can be found in the root directory.
Activate it by running:
conda env create -f env522car.yaml
conda activate 522env
You can also manually install the following packages:
- Python 3.10.6 and Python Packages:
- docopt-ng 0.8.1
- requests 2.28.1
- pandas 1.5.1
- altair 4.2.0
- altair_data_server 0.4.1
- joblib 1.2.0
- matplotlib
- jsonschema 4.16
- selenium <4.3.0
- docopt-ng 0.8.1
- vl-convert-python 0.5.0
- eli5 0.13.0
- shap 0.41.0
- imbalanced-learn 0.9.1
- lightgbm 3.3.3
- pip 22.3.1
- jinja2 3.1.2
- ipykernel 6.18.0
- altair_saver 0.5.0
- python-graphviz 0.20.1
- graphviz 6.0.1
- scikit-learn 1.1.3
- R 4.2.1 and R packages:
- tidyverse. 1.3.2
- knitr 1.40
- kableExtra 1.3.4
- pandoc 0.1.0
- GNU make 4.2.1
click the image for better visability
The Car Popularity Predictor materials here are licensed under Creative Commons Attribution 4.0 Canada License (CC BY 4.0 CA).