Dual-View Desynchronization Hypergraph Learning for Dynamic Hyperedge Prediction

• Published in: IEEE transactions on knowledge and data engineering 2024-11, p.1-16
• Main Authors: Wang, Zhihui, Chen, Jianrui, Shao, Zhongshi, Wang, Zhen
• Format: Article
• Language: English
• Subjects: Aggregates; Analytical models; Biological neural networks; Contrastive learning; Decoding; Desynchronization learning; dual-view learning; dynamic hypergraph learning; Feature extraction; graph convolutional network; hyperedge prediction; Predictive models; Representation learning; Training; Transforms
• ISSN: 1041-4347; 1558-2191
• DOI: 10.1109/TKDE.2024.3509024

Hyperedges, as extensions of pairwise edges, can characterize higher-order relations among multiple individuals. Due to the necessity of hypergraph detection in practical systems, hyperedge prediction has become a frontier problem in complex networks. However, previous hyperedge prediction models encounter three challenges: (i) failing to predict dynamic and arbitrary-order hyperedges simultaneously, (ii) confusing higher-order and lower-order features together to propagate neighborhood information, and (iii) lacking the capability to learn physical evolution laws, which lead to poor performance of the models. To tackle these challenges, we propose D ^{3} HP, a D ual-view D esynchronization hypergraph learning for arbitrary-order D ynamic H yperedge P rediction. Specifically, D ^{3} HP extracts the dynamic higher-order and lower-order features of hyperedges separately through an elastic hypergraph neural network (EHGNN) and an alternate desynchronization graph convolutional network (ADGCN) at each time snapshot. EHGNN is designed to incrementally mine the implicit higher-order relations and propagate neighborhood information. Moreover, ADGCN aims to combine GCN with desynchronization learining to learn the physical evolution of lower-order relations and alleviate the over-smoothing problem. Further, we improve the prediction performance of the model by rationally fusing the features learned from the dual views. Extensive experiments on 8 dynamic higher-order networks demonstrate that D ^{3} HP outperforms 14 state-of-the-art baselines.