GENIE Graph Anomaly Detection
Overview
Developed an end-to-end graph autoencoder pipeline for detecting anomalous particle-collision events in LHCO-style datasets. The system converts raw HDF5 events into subjet graphs, learns background structure without anomaly labels, and assigns event-level scores from graph reconstruction errors.
Key result: Adjacent-subjet regularization improved BB1 AUROC from 0.891 to 0.923, MaxSIC from 2.40 to 3.87, and signal efficiency at 1% background efficiency from 0.236 to 0.355.
End-to-End Pipeline
1. Event Processing
Reads LHCO HDF5 events, clusters anti-kT jets, applies a 1.2 TeV leading-jet threshold, and retains two selected jets per event.
2. Graph Construction
Reclusters each jet into 30 exclusive-kT subjets and builds unique-6 graphs with log ฮR, kT, and momentum-fraction edge features.
3. Graph Autoencoder
Jointly reconstructs node and edge information with EdgeConv message passing; alternative GCN, GraphSAGE, GATv2, GIN, and Transformer backbones support ablations.
4. Event-Level Evaluation
Aggregates selected-jet reconstruction errors into anomaly scores and evaluates AUROC, MaxSIC, and signal efficiency on LHCO and BB1.
Topology-Aware Learning
The adjacent-subjet objective preserves local graph structure in the learned latent space. It adds a distance penalty between representations of connected subjets:
L = Lrecon + ฮปR(E, z), where R(E, z) = (1/N) ฮฃ(i,j)โE โzi โ zjโ.
The experiments compare physical graph edges with adjacency rebuilt from detached latent representations at different encoder depths. This isolates whether the gain comes from input topology or learned latent geometry.
Results
LHCO
AUROC 0.926 vs. 0.910 baseline
MaxSIC 2.94 vs. 2.32 baseline
BB1 Transfer Evaluation
AUROC 0.923 vs. 0.891 baseline
MaxSIC 3.87 vs. 2.40 baseline
Evaluation discipline: The baseline uses the final checkpoint from a fixed 50-epoch unsupervised schedule; the regularized result uses the final checkpoint after five additional fine-tuning epochs. Signal labels are reserved for evaluation and diagnostic plots rather than checkpoint selection.
Engineering Contributions
- Implemented a reproducible five-stage pipeline covering preprocessing, graph building, event splitting, training, and external-dataset evaluation.
- Designed sharded PyTorch Geometric datasets with caching and GPU prefetch for large collider-event collections.
- Added six graph autoencoder families and configurable message-passing backbones behind a validated model factory.
- Built checkpoint contracts, reusable event manifests, topology ablations, and mJJ-window training for controlled experiments.