Records
1.1M+
collision records with road topology
Back to projects
Independent / Applied ML Project
Graph-based traffic collision risk modeling with leakage-aware validation.
This project modeled collision risk over 1.1M+ collision records and road-network topology, representing intersections, road segments, and neighborhoods with spatial, temporal, traffic, structural, weather, and historical-risk covariates.
Key Outcomes
Graph structure improved risk ranking under stricter validation.
Ranking
+23.4
hotspot-ranking AUROC points over tabular baselines
Recall
+31%
top-decile recall improvement
Project Breakdown
Problem, method, system, validation, results, reliability, and research value.
Problem
Spatial prediction can leak information through time and geography.
- Nearby road segments and future collision patterns can make naive validation look better than deployable performance.
- The project needed to separate graph-structure signal from temporal leakage and spatial autocorrelation artifacts.
Method
Collision risk was formulated as heterogeneous spatiotemporal graph prediction.
- Nodes and edges represented intersections, road segments, and neighborhoods with spatial, temporal, traffic, structural, weather, and historical-risk covariates.
- Compared message passing against temporal, tabular, kernel, XGBoost, and geospatial baselines.
System / Stack
The pipeline joined geospatial processing with graph ML.
- Used Python, PyTorch Geometric, NetworkX, GeoPandas, road-network topology, SQL, XGBoost, spatial statistics, and calibration tooling.
- Built feature construction, graph assembly, model comparison, calibration, and risk-map outputs.
Validation Methodology
Evaluation used temporal and geographic defenses against leakage.
- Used temporal cutoffs, geographic buffer zones, held-out corridors, future-information audits, and spatial autocorrelation diagnostics.
- Compared GCN, GraphSAGE, GATv2, temporal aggregation, XGBoost, kernel, tabular, and non-graph geospatial models.
Results
Topology improved hotspot ranking.
- Improved hotspot-ranking AUROC by 23.4 points and top-decile recall by 31% over tabular baselines.
- Connectivity, neighborhood propagation, centrality, temporal exposure windows, and message passing contributed to the lift.
Failure Modes / Reliability Checks
The model was stress-tested for spatial shortcuts.
- Added calibrated risk maps, ablations, counterfactual edge removal, conformal-style risk sets, reliability curves, ranking-stability checks, and future-information audits.
Why It Matters for Research
The project connects graph learning to scientific validation discipline.
- It asks whether relational structure improves risk modeling after the validation protocol removes shortcuts that would not survive deployment.