How Much Does Fusion Design Matter? Partition, Seed, and Cluster-Aware Evaluation of Hybrid LSTM–GAT Traffic Anomaly Detectors

Architectural assessments of traffic anomaly detectors can be sensitive to data partitioning, random training seeds, and dependence among overlapping windows. We evaluate nine encoder configurations, four fusion operators, and two graph constructions using physics-informed synthetic trajectories generated on an OpenStreetMap-derived road network. Each fold–seed threshold is selected on validation data and applied to its corresponding test set. Trip-level cluster bootstrap and permutation tests account for dependence among windows, while corrected resampled tests approximate the dependence induced by overlapping training sets. On a fixed partition of 14 test trips, no comparison establishes superiority after Holm correction. Across five-fold cross-validation with five seeds, mean F1 scores range from 0.8558 to 0.8839 across fusion operators, with gating and cross-attention producing occasional degenerate runs. Across 500 runs covering repeated grouped partitions and both graph constructions, no pairwise architectural comparison remains significant after correction. With run-specific thresholds, subtype analyses do not support the abnormal-dwell improvement that appears under pooled thresholds. Classical baselines remain competitive, but their apparent advantage in single-seed pooled tests is not corroborated by run-level inference. Overall, the results demonstrate that architectural conclusions are sensitive to threshold selection, the number of training seeds, and the choice of statistical unit of analysis.