v1.0.0: Elliptic++ Fraud Detection - When Do GNNs Add Value? 🎯 Release Summary This release marks the completion of a comprehensive comparative study investigating when Graph Neural Networks provide marginal value over tabular ML models for fraud detection on the Elliptic++ Bitcoin transaction dataset. Main Finding: Features already encode graph aggregations — GNNs redundant unless features are raw. 🔬 Key Research Contributions 1. Feature Dominance Hypothesis — CONFIRMED ✅ Through ablation experiments (M7), we confirmed that tabular features AF94–AF182 already encode neighbor-aggregated information: | Model | Config | PR-AUC | Δ vs Full | Interpretation | |-------|--------|--------|-----------|----------------| | XGBoost | Full (AF1–182) | 0.669 | — | Baseline | | XGBoost | Local only (AF1–93) | 0.648 | −3% | Barely affected | | GraphSAGE | Full (AF1–182) | 0.448 | — | Redundant encoding | | GraphSAGE | Local only (AF1–93) | 0.556 | +24% | GNN unlocked! | Evidence: Neighbor averages correlate r=0.74–0.89 with aggregate features Manual graph metrics correlate r=0.63–0.65 with aggregates AF94–AF182 are literally pre-computed neighbor aggregations 2. Interpretability Analysis (M8) XGBoost (full): Heavily relies on aggregate features (SHAP analysis) GraphSAGE (local-only): Learns from graph structure via message passing Models learn different representations: pre-computed features vs dynamic aggregation 3. Temporal Robustness (M9) XGBoost: Stable 0.67–0.78 PR-AUC across time shifts GraphSAGE (local): Improves 0.41 → 0.56 with earlier training windows (+35%) Finding: GNNs handle temporal drift better when trained on raw features 📊 Complete Results Model Performance Rankings (PR-AUC): | Rank | Model | Type | PR-AUC | ROC-AUC | F1 | Recall@1% | |------|-------|------|--------|---------|----|-----------| | 🥇 1 | XGBoost | Tabular | 0.669 | 0.888 | 0.699 | 0.175 | | 🥈 2 | Random Forest | Tabular | 0.658 | 0.877 | 0.694 | 0.175 | | 🥉 3 | GraphSAGE | GNN | 0.448 | 0.821 | 0.453 | 0.148 | | 4 | MLP | Tabular | 0.364 | 0.830 | 0.486 | 0.094 | | 5 | GCN | GNN | 0.198 | 0.763 | 0.249 | 0.061 | | 6 | GAT | GNN | 0.184 | 0.794 | 0.290 | 0.013 | | 7 | Logistic Regression | Tabular | 0.164 | 0.824 | 0.256 | 0.005 | 🎓 Milestones Completed (10/10) ✅ M1: Repository scaffold & infrastructure ✅ M2: Dataset loader with temporal splits ✅ M3: GCN baseline implementation ✅ M4: GraphSAGE & GAT models ✅ M5: Tabular baselines (LR, RF, XGBoost, MLP) ✅ M6: Documentation & comparative analysis ✅ M7: Causality & Feature Dominance — HYPOTHESIS CONFIRMED ✅ M8: Interpretability (SHAP + GNN saliency) ✅ M9: Temporal Robustness Study ✅ M10: Final polish & release preparation 📁 What's Included Documentation PROJECT_REPORT.md — Publication-style comprehensive report (13KB) README.md — Complete project overview with findings PROJECT_SUMMARY.md — Detailed narrative and evidence docs/M7_RESULTS.md — Feature dominance ablation results docs/M8_INTERPRETABILITY.md — SHAP + GNN saliency analysis docs/M9_TEMPORAL.md — Temporal robustness findings LICENSE — MIT License Code & Notebooks 8 reproducible notebooks (notebooks/03-08_*.ipynb) 12+ training scripts (scripts/run_m*.py) Model implementations (GCN, GraphSAGE, GAT) Complete data pipeline Artifacts 25+ result files in reports/ M7 ablation experiments (7 CSV/JSON files) M8 interpretability (3 CSV/JSON + plots) M9 temporal robustness (1 CSV file) All baseline model metrics 💡 Key Takeaways Graph structure is valuable — but dataset features already captured it through pre-computed aggregations. For Practitioners: ✅ Use XGBoost when features include graph aggregations (fast, interpretable, CPU-friendly) ✅ Use GNNs when features are raw and graph structure is critical ✅ Always check feature-structure redundancy before selecting models For Researchers: Feature engineering quality matters more than model architecture Ablation studies reveal hidden feature-structure redundancies Correlation analysis essential for understanding feature provenance Temporal robustness tests reveal model generalization characteristics 🚀 Getting Started # Clone repository git clone https://github.com/BhaveshBytess/FRAUD-DETECTION-GNN.git cd FRAUD-DETECTION-GNN # Install dependencies pip install -r requirements.txt # Explore results jupyter notebook notebooks/ # Read full report cat PROJECT_REPORT.md 📖 Citation If you use this work, please cite: @misc{elliptic-gnn-2025, title={Elliptic++ Fraud Detection: When Do Graph Neural Networks Add Value?}, author={Bhavesh Bytes}, year={2025}, url={https://github.com/BhaveshBytess/FRAUD-DETECTION-GNN}, note={v1.0.0} } 🎉 Acknowledgments Elliptic++ dataset providers PyTorch Geometric community Kaggle for GPU resources Project Status: ✅ Complete Release Date: 2025-11-08 License: MIT ⭐ If you find this project useful, please star the repository!