joelinator/resp_ai_project

Project for the 2026 Responsible AI course at AIMS South Africa # Final Project Pipeline This repository contains a complete end-to-end implementation of my final project of the responsible AI course at AIMS south Africa ## Run Use the requested virtual environment Python: ```bash /home/joelinator/env/bin/python run_project.py ``` Reframed 5x simulation run (used by latest report/presentation): ```bash /home/joelinator/env/bin/python run_project.py --simulation-timestep-multiplier 5 --results-dir results_reframed_proficiency_5x ``` ## What the script does 1. Downloads ASSISTments 2009-2010 data if missing. 2. Preprocesses and creates deterministic synthetic gender. 3. Trains a calibrated KT model (RandomForest + Platt scaling). 4. Simulates heterogeneous teachers. 5. Evaluates: - baseline confidence-threshold policy, - proposed LinUCB routing with fairness guardrail. 6. Logs metrics, fairness tables, policy traces, and plots to `results/`. 7. Generates transparency artifacts (global + local SHAP plots). 8. Copies NeurIPS style/checklist assets into `paper/`. ## Outputs - `results/metrics_summary.csv` - `results/fairness_metrics.csv` - `results/model_metrics.json` - `results/risk_coverage_*.png` - `results/shap_*.png` - `results/baseline_policy_logs.csv` - `results/proposed_policy_logs.csv` - `results/guardrail_history.csv` - `results/analysis.md` - `paper/paper.tex` and `paper/references.bib` ## Overleaf (recommended) For easiest compilation on Overleaf, upload the whole repository and set the main file to: - `main.tex` Why: - `main.tex` is root-based (no parent `..` image paths), - it references all needed assets directly from `results/`, `images/`, and `paper/`, - it is robust to either flowchart filename: - `images/proposed_vs_baseline_flowchart.png`, or - `images/proposed_vs_baseline_flowchart.txt.png`.