CF-TGNN: Counterfactual Explainable Temporal Graph Neural Networks for Supply Chain Disruption Prediction

Disruptions in multi-tier supply chains travel from upstream suppliers to downstream customers, and a risk model helps planners only if it can say what would have prevented a predicted disruption. Temporal graph neural networks with supplier attention forecast such ripple effects well, but attention weights are normalized over a node’s suppliers: they always place blame upstream, cannot state that no supplier is responsible, and do not describe an intervention. We present CF-TGNN, a counterfactual explainable temporal graph neural network. Its backbone adopts a directed, dependency-weighted temporal attention operator and adds a region-context encoder that acts as a proxy for confounding regional shocks. Its Counterfactual Restoration Explainer (CRE) evaluates supplier-restoration interventions with do-operator graph surgery and returns a but-for estimate of the probability that upstream disruptions were necessary, a restoration Shapley decomposition of the predicted risk, an exact minimum-cardinality restoration plan, and an origin trace that follows causes backward in time. To test explanations against true counterfactuals, we turn a published cascade simulator into a structural causal model with logged exogenous noise, so that every what-if query can be answered by replay. Over five seeds, CF-TGNN raises onset AUC from 0.765 to 0.786; CRE identifies cascade origins with 44.2% accuracy versus 33.8% for layer-wise attention chaining; and graph surgery reduces the error of but-for estimates by 22%. A case study on the public SupplyGraph dataset shows that CRE returns sparse restoration plans on real production records and makes disagreements between backbones explicit, although the region context does not improve prediction on these data.