How vjs5/fair-active-learning-ai4h works
This repository is a research implementation of fair active learning for credit-risk classification. Its runtime loop loads the credit dataset, repeatedly selects and labels candidates using one of four strategies, refits a Bayesian logistic model, and records representation, predictive, fairness, calibration, and uncertainty metrics; visualization then summarizes strategy results. The excerpts support the model and reporting capabilities, but do not show the loop’s invocation or the plotting orchestration, so those connections are omitted rather than inferred.
Main entry points: model.py Active-learning loop · visualize.py Diagnostic plots
Components 11
Active learning
- Active-learning loop
src/model.py - FAL strategy selector
src/model.py - Fairness-gain estimate
src/model.py
Bayesian modeling
- Bayesian logistic regression
src/model.py - Predictive uncertainty
src/model.py
Evaluation and reporting
- Fairness metrics
src/model.py - Calibration metrics
src/model.py - Run history
src/model.py - Diagnostic plots
src/visualize.py - Metrics summary
src/visualize.py
Other components
- Research user
Connections 12
- Research user → Active-learning loop: runs experiment
- Active-learning loop → FAL strategy selector: requests selection
src/model.py - FAL strategy selector → Predictive uncertainty: scores uncertainty
src/model.py - FAL strategy selector → Fairness-gain estimate: scores fairness gain
src/model.py - Fairness-gain estimate → Bayesian logistic regression: samples posterior
src/model.py - Active-learning loop → Bayesian logistic regression: refits model
src/model.py - Bayesian logistic regression → Predictive uncertainty: predicts entropy
src/model.py - Active-learning loop → Fairness metrics: evaluates predictions
src/model.py - Active-learning loop → Calibration metrics: evaluates probabilities
src/model.py - Active-learning loop → Run history: records metrics
src/model.py - Run history → Diagnostic plots: plots results
src/visualize.py - Run history → Metrics summary: summarizes results
src/visualize.py