Bias Detection & Interpretability

Overview

Category: Bias Detection

Modular Components: - Fairness Metrics Evaluator - Model Explainer (SHAP/LIME) - Feature Attribution Visualizer

Use Cases

• Do public service recommendation models (e.g., bus stop placement, streetlight allocation) tend to ignore low-density or low-income areas?
• Why citizens’ application for housing subsidies or public services was deprioritized?
• Do predictive maintenance systems for infrastructure (e.g., water leaks, power outages) prioritize certain zones? Is this optimization fair?
• In citizen feedback systems (e.g., 311 complaints), are certain types of reports more likely to trigger response recommendations than others? Why?

Code Example: Credit Decision Bias Auditing

Content: - Analyze credit data using LLM and interpretable ML - Detect bias in approval logic (e.g., income, gender) - Apply SHAP and counterfactual fairness methods

Datasets: - German Dataset (credit data) from AIF Fairness 360 - COMPAS dataset (pre-prepared from https://www.kaggle.com/datasets/danofer/compass)

Required Packages: Fairlearn, SHAP, pandas, scikit-learn, transformers