How ru-insane/compas-fairness-privacy-tradeoff-analysis works
Major Components
• Data Layer – Represented by the “data” folder, which includes CSV files for training, testing, baseline results, logistic regression predictions, ROC curves, etc.
• Baseline Processing & Modeling – Found in “main/baseline”. This includes notebooks for data pre-processing, exploratory data analysis (EDA), and model training (both logistic regression and multi-layer perceptron).
• Fairness Analysis – Under “main/fairness”, which includes notebooks for fairness preprocessing, in-processing, and post-processing. This clearly delineates the component that adjusts or analyzes fairness measures.
• Privacy Analysis – Available in “main/privacy”, focusing on differential privacy methods, showing implementation of privacy adjustments to the modeling process.
• Merged Analysis – In “main/merged”, where fairness and privacy considerations are combined. This serves as an integration layer to evaluate tradeoffs.
• Model Artifact – The “model” folder, which contains the saved multi-layer perceptron model file (mlp_model.pth).
• Supplementary Documentation – Including “documentation” and “Research References” folders, they give context and the research basis for the experiments.
Relationships and Interactions
• Decision Branching: From the baseline component, the system diverges into separate paths — one focusing on fairness-related processing (pre-, in-, and post-processing), and another on applying privacy techniques.
• Integration: Both the fairness and privacy pipelines then converge into the “merged” component where tradeoffs are jointly analyzed and compared.
• Output and Model Persistence: The trained models (especially the MLP model) are saved in the “model” folder, which may then be used for further evaluation or deployment (if extended).
Architectural Patterns and Design Principles
• Modular Pipeline Structure: The clear separation into baseline, fairness, privacy, and merged experiments demonstrates a modular design that isolates different concerns (data processing, fairness adjustments, privacy controls, and integration analysis).
• Experimentation and Iterative Refinement: Notebooks indicate an iterative, exploratory process, and the division of notebooks signifies discrete experimental steps.
• Data-Driven Flow: The system is driven by structured inputs (CSV data) and has output evaluations (results, ROC curves, logs).
Relevant Technologies and Tools
• Jupyter Notebooks as the primary interface for experimentation.
• Data files in CSV format, indicating traditional data engineering methods.
• Python (implied by ".ipynb" notebooks and requirement.txt) along with machine learning frameworks (as seen in “mlp_model.pth” which likely results from a PyTorch-based implementation).
• Documentation and reference materials in standard document formats (DOCX, PDF, XLSX) that support comprehensive background research.