A Hypergraph Neural Network Framework for Resilience Inference in Complex Systems: Evidence from Urban Rail Transit

Abstract

Resilience characterizes the ability of a complex network to maintain core functions under disturbances. Traditional analytical approaches often rely on predefined governing equations and may have difficulty representing higher-order interactions. To address these limitations, this study proposes ResInf-HGNN, a resilience inference framework that combines a Transformer temporal encoder with hypergraph convolution. Hyperedges represent line-level group relationships among stations, while high-dimensional spatiotemporal states are projected into a one-dimensional k-space. Once trained, the model infers resilience from topology and early trajectories without requiring explicit governing equations at inference time. Training and evaluation data for the Nanjing rail transit network are generated by a physically constrained coupled physical-dispatch dynamics model that combines low-order diffusion with line-level dispatch cascades. On the 1:4 imbalanced test set, ResInf-HGNN achieves an F1-score of 0.9362, AUC-ROC of 0.9965, and precision of 0.8800. Ablation results show that removing the Transformer and attention readout reduces the F1-score to 0.9167 and 0.8718, respectively. Under node perturbations, targeted node attacks generally cause greater structural damage, whereas random failures yield the lowest minimum robustness R_ob among the compared attack modes. With progressive removal of line relations from the hypergraph, normalized hypergraph connectivity falls to zero while normalized R_sys rises to 1.0560; this increase reflects reduced simulated exposure to the higher-order dispatch-cascade mechanism and should not be interpreted as improved real-world operational resilience. Overall, ResInf-HGNN provides a data-driven framework for resilience inference and dual-perspective analysis of rail transit networks.