Physics-Guided Hypergraph Contrastive Learning for Dynamic Hyperedge Prediction

With the increasing magnitude and complexity of data, the importance of higher-order networks is increasingly prominent. Dynamic hyperedge prediction reveals potential higher-order patterns with time evolution in networks, thus providing beneficial insights for decision making. Nevertheless, most existing neural network-based hyperedge prediction models are limited to static hypergraphs. Furthermore, previous efforts on hypergraph contrastive learning involve augmentation strategies, with insufficient consideration of the higher-order and lower-order views carried by the hypergraph itself. To address the above issues, we propose PCL-HP, a physics-guided hypergraph contrastive learning framework for dynamic hyperedge prediction. Specifically, we simply distinguish higher-order and lower-order views of the hypergraph to perform dynamic hypergraph contrastive learning and obtain abstract and concrete feature information, respectively. For lower-order views, we propose a physics-guided desynchronization mechanism to effectively guide the encoder to fuse the physical information during feature propagation, thus alleviating the problem of feature over-smoothing. Additionally, residual loss is introduced into the optimization process to incrementally quantify the loss at different stages to enhance the learning capability of the model. Extensive experiments on 10 dynamic higher-order datasets indicate that PCL-HP outperforms state-of-the-art baselines.