How jma712/gear works
Project Type and Purpose
• Type: Research‑oriented ML pipeline / open‑source tool for fair graph node embedding.
Main Components
A. Data Layer
– Raw datasets (dataset/bail, dataset/credit, dataset/synthetic.mat)
– CSV/edge‑list files and dataset readmes
B. Preprocessing
– Preprocessing.py: cleans, normalizes node features and sensitive attributes
– src/graphFair_subgraph/cf: load true counterfactuals
D. Model Definition
– models.py: GEAR architecture (GNN encoder, fairness loss)
– CFGT.py & CFDA.py: implementations of counterfactual generation/training strategies
E. Training & Evaluation Orchestration
– utils.py & utils_mp.py: multiprocessing, data loaders, metric computations
F. Persistence & Outputs
– models_save/: trained model checkpoints
– logs/metrics (you may add if desired)
Relationships & Data Flow
1. Raw Data → Preprocessing
4. Training Iterations ↔ Models & Utils
5. Model Checkpoints → models_save/
6. Evaluation → metrics printed or persisted
Architectural Patterns & Design Principles
• Modular separation: each major step lives in its own script/module
• Reproducibility: CLI interface in main.py captures experiment_type and hyperparameters
Technologies & Frameworks
• Python 3.6
• PyTorch 1.6.0 + torch‑geometric 1.3.0 (GNN layers, data loaders)
• scikit‑learn (metrics, preprocessing)
• numpy (matrix ops)
• multiprocessing (utils_mp.py)