Explainable Relation-Heterogeneous Temporal Graph Neural Networks for Supply Chain Disruption Onset Prediction

Abstract

Supply-chain disruption prediction is structurally different from ordinary time-series forecasting because risk is transmitted through typed dependencies such as sourcing, transportation, distribution, and alternate-supply relations. We present RHET-GNN, a relation-heterogeneous temporal graph neural network that couples a per-entity GRU encoder with directed, relation-specific, dependency-weighted graph attention. The same normalized attention coefficients provide an intrinsic explanation over incoming typed edges, allowing a predicted onset to be decomposed by supplier and relation without invoking a separate post-hoc explainer. Because publicly available supply-chain benchmarks such as SupplyGraph support planning tasks but do not provide event-level ground-truth disruption-propagation paths, we evaluate on a fully reproducible four-type, four-relation synthetic testbed whose causal edge contributions are logged by construction. No experimental values are manually specified: all reported metrics are produced by the accompanying code. Across three independently generated seeds, RHET-GNN obtains onset AUC 0.723±0.010, AP 0.252±0.006, and F1 0.325±0.007. Relative to a temporal-only GRU, AP improves by 79.0%. Relative to a relation-agnostic temporal graph-attention baseline, predictive AP is similar (+0.9% relative), while explanation precision@1 improves by 4.4 percentage points and NDCG by 2.5 points. These results indicate that graph structure drives most of the predictive gain, whereas explicit relation typing is especially valuable for attribution quality.