Deep Graph Anomaly Detection
Overview
Completed an end-to-end research and engineering pipeline for detecting anomalous particle-collision events with graph learning. The project converts raw collider events into subjet graphs, benchmarks graph autoencoders and self-discriminative models, and couples learned physics representations to a 50-model CWoLa gradient-boosting ensemble.
Key results: Physics-aware NetGe features reached 0.9542 AUROC / 39.92 maxSIC on LHCO. Dual-view InfoNCE then improved mean AUROC across five signals by 0.52 percentage points and raised mean maxSIC from 43.83 to 45.04.
Research Pipeline
1. Events to Graphs
Clusters anti-kT jets, selects the two leading jets, reclusters up to 30 subjets, and builds Unique-6 graphs with physical node and edge features.
2. Graph Model Research
Benchmarks EdgeConv, GLADC, NetGe, and self-discriminative models across LHCO, BB1, and four Kitchen Sink signals.
3. Physics-Aware Features
Trains autoencoders to reconstruct graphs and predict mass, N-subjettiness, energy-flow polynomials, ฯ ratios, and Soft Drop quantities.
4. CWoLa Tree Readout
Freezes the learned representations, trains 50-model SR/SB gradient-boosting ensembles, and reports multi-seed AUROC and maxSIC.
Physics-Aware Representation Learning
The NetGe autoencoder reconstructs node and physical-edge features while auxiliary heads predict jet mass, N-subjettiness, energy-flow polynomials, and ฯ ratios. Graph-derived mass restores the absolute momentum scale removed by relative-pT inputs.
80 tree features: 2 graph-mass features + 54 ฯ predictions + 16 EFP predictions + 2 reconstruction-error features + 6 log ฯ-ratio predictions.
After training on sideband events, the autoencoder is frozen and its predictions become inputs to the CWoLa tree ensemble. The learned features, rather than reconstruction error alone, produce the final event-level anomaly score.
Dual-View Contrastive Learning
NetGe and a pair-attention autoencoder provide complementary sparse-graph and all-pairs views of the same event. Each model first trains for 50 epochs; they then train jointly for another 50 epochs with InfoNCE over their 70 ฯ/EFP predictions.
L = LNetGe + LAttention + 0.03 LInfoNCE. Predictions from the same event form positive pairs, while other events in the batch are negatives.
The contrastive weight is calibrated from gradient norms, and both encoders receive gradients. The final tree ensemble uses 154 features spanning both models' predictions, reconstruction errors, Soft Drop quantities, and graph-derived masses.
Results
Physics-Aware NetGe + Trees
0.9542 AUROC
39.92 maxSIC on LHCO
Dual AE + InfoNCE
0.9563 ยฑ 0.0026 AUROC
44.22 ยฑ 2.97 maxSIC on LHCO
Five-Signal Average
+0.52 percentage points AUROC
43.83 โ 45.04 maxSIC
Unsupervised NetGe
0.9239 / 2.966 on LHCO
0.9017 / 3.519 on full-range BB1
Evaluation discipline: Dual-AE results are mean ยฑ sample standard deviation across seeds 42, 43, and 44 with fixed event splits and tree seeds. InfoNCE improves mean AUROC on all five signals and maxSIC on four.
Engineering Contributions
- Built reproducible preprocessing, graph construction, SR/SB splitting, training, inference, and evaluation workflows.
- Implemented and benchmarked EdgeConv, GLADC, NetGe, and self-discriminative graph models across six collider datasets.
- Added physics-prediction heads, pair-attention encoding, dual-AE contrastive training, and 50-model tree readout.
- Established fixed-split, multi-seed evaluation with documented commands, metrics, figures, and ablation results.