/COMPAS-Recidivism-Bias-Analysis

Primary LanguageJupyter Notebook

COMPAS Recidivism Bias Analysis

Overview

This repository contains a machine learning project focused on analyzing biases in the COMPAS (Correctional Offender Management Profiling for Alternative Sanctions) dataset. The goal is to identify and assess potential biases in recidivism predictions across different racial groups using a decision tree classifier.

Project Description

The COMPAS dataset is widely used to assess the likelihood of a convict's recidivism. However, it has been subject to scrutiny and debate due to potential biases against certain demographic groups, particularly in terms of race. This project aims to explore these potential biases by applying a decision tree model and evaluating its performance across different racial subgroups.

Repository Structure

- data/                        # Folder containing the COMPAS dataset and any other data files
- notebooks/                   # Jupyter notebooks with detailed analysis
- src/                         # Source code for analysis, including data preprocessing and modeling
- results/                     # Generated output files, figures, and confusion matrices
- README.md                    # Description and summary of the project
- requirements.txt             # List of dependencies for reproducing the analysis

Key Findings

  • The decision tree model showed varying degrees of performance across different racial groups.
  • Significant False Positives and False Negatives were observed, particularly between classes 1 and -1.
  • The model’s predictions indicated potential biases, necessitating further investigation and refinement.

Methodology

  1. Data Preprocessing: The dataset was cleaned and prepared, involving handling missing values, encoding categorical variables, and feature engineering.
  2. Model Training and Evaluation: A decision tree classifier was trained and evaluated, with performance metrics calculated for the overall model and for each racial subgroup.
  3. Bias Analysis: Confusion matrices were generated for each racial subgroup to assess the model's fairness and accuracy.

Technologies Used

  • Python
  • Pandas, NumPy
  • Scikit-learn
  • Jupyter Notebook

Running the Project

To replicate this analysis, clone the repository, install the dependencies listed in requirements.txt, and run the Jupyter notebooks in the notebooks/ directory.

git clone https://github.com/elshadaik/COMPAS-Recidivism-Bias-Analysis.git
pip install -r requirements.txt

Contributing

Contributions to this project are welcome. Please fork the repository and submit a pull request with your changes.

License

This project is open-sourced under the MIT License.

Acknowledgements

Special thanks to all contributors and researchers who have worked on analyzing and highlighting issues in the COMPAS dataset.