Dual Spatial-Temporal Attribution: Architecture-Aligned Post-Hoc Explainability for Recurrent Graph Anomaly Detection
X-AddGraph introduces the first post-hoc explainability framework for AddGraph, a foundational GCN+GRU model for edge-level anomaly detection in dynamic graphs, addressing a critical opacity gap in regulated AI systems. The Dual Spatial-Temporal Attribution (DSTA) mechanism aligns three attribution components with AddGraph's architecture: gradient-based spatial relevance, zero-cost short-term temporal attribution via existing attention weights, and gradient rollback through recurrent hidden stat
Analysis
TL;DR
- X-AddGraph introduces the first post-hoc explainability framework for AddGraph, a foundational GCN+GRU model for edge-level anomaly detection in dynamic graphs, addressing a critical opacity gap in regulated AI systems.
- The Dual Spatial-Temporal Attribution (DSTA) mechanism aligns three attribution components with AddGraph's architecture: gradient-based spatial relevance, zero-cost short-term temporal attribution via existing attention weights, and gradient rollback through recurrent hidden states for long-term temporal attribution.
- The detector remains completely frozen during explanation generation, preserving detection performance exactly (Delta AUC = 0 to ten decimal places), with the trained baseline achieving 0.8705 average per-snapshot AUC on UCI Message, exceeding the originally published result.
- Long-term attribution identifies historical snapshots with significantly more counterfactual signal than random selection (0.127 vs. 0.074), a capability unattainable by spatially-blind explainers, validated across confident true positives, low-confidence true positives, false positives, and random edge samples.
- Full implementation is released for reproducibility, enabling auditable and trustworthy deployment of graph anomaly detectors in cooperative, regulated information systems.
Why It Matters
This work bridges a critical trust gap in dynamic graph anomaly detection: high-performing deep learning models have been deployed without any explainability, making them unsuitable for regulated environments where automated decisions must be auditable. By providing architecture-aligned, strictly post-hoc explanations without compromising detection performance, X-AddGraph enables practitioners to deploy graph-based anomaly detectors in production settings where accountability and transparency are mandatory, such as financial fraud detection, cybersecurity, and network monitoring.
Technical Details
- Architecture-aligned DSTA mechanism: Three attribution components map directly to AddGraph's modules—a gradient-based relevance attribution over the current adjacency matrix (spatial), direct extraction of contextual attention weights computed during inference (short-term temporal, zero additional computational cost), and gradient rollback through GRU recurrent hidden states (long-term temporal).
- Post-hoc and non-invasive design: The detector is entirely frozen; no retraining or architectural modification is required. Performance preservation is empirically verified to ten decimal places (Delta AUC = 0).
- Benchmark performance: On the UCI Message dynamic graph benchmark, the retrained AddGraph baseline achieves 0.8705 average per-snapshot AUC, surpassing the originally published result, while X-AddGraph reproduces every detection score identically and adds explanatory outputs.
- Counterfactual signal evaluation: Across four edge populations (confident TP, low-confidence TP, FP, random), long-term attribution selects historical snapshots carrying meaningfully more counterfactual signal (0.127) compared to random selection (0.074), demonstrating that temporal attribution captures genuinely informative historical context.
- Reproducibility: The complete implementation is publicly released, supporting full reproducibility and enabling downstream research on explainable graph anomaly detection.
Industry Insight
- The architecture-aligned approach—mapping each attribution component to a specific model module—offers a reusable blueprint for adding explainability to other recurrent graph models without sacrificing performance, suggesting that post-hoc explainability can be systematically integrated rather than bolted on ad hoc.
- The finding that long-term temporal attribution captures counterfactual signal unavailable to spatial-only explainers highlights the importance of temporal interpretability in dynamic systems; practitioners deploying graph anomaly detectors should prioritize methods that expose historical context, not just current structure.
- As regulatory frameworks increasingly mandate auditability for automated decisions (e.g., EU AI Act), this work demonstrates that high-accuracy deep learning models for graph analytics can meet compliance requirements without performance trade-offs, making explainable graph AI a near-term practical necessity rather than a theoretical concern.
Disclaimer: The above content is generated by AI and is for reference only.