Knowledge Hypergraph Link Prediction Based on Attention and Relational Contrast Learning

Knowledge Hypergraph (KHG) is a specialized form of knowledge graph that uses a hypergraph structure to represent n-ary relationships. Link prediction in KHGs aims to infer missing entities and relations from existing ones, and is an essential task for many knowledge-driven applications. However, current neural approaches often struggle to model interfact dependencies and maintain semantic consistency across complex relational patterns. To address these challenges, we propose a hypergraph link prediction method based on attention and convolutional neural network, called HPAC. Firstly, our refined attention module assigns distinct weights to nodes and hyperedges, enabling effective aggregation of local and global contexts to capture intricate relational dependencies. Secondly, inspired by contrastive learning, we propose a context-level contrastive model to capture the latent semantics among related entities and refine their embeddings. By constructing multiple contextual views through cross-perspective contrastive strategies, HPAC enriches node and edge representations with diverse semantic cues, thereby boosting its discriminative power. Finally, we apply regularization techniques to mitigate neuron overreliance, enhancing prediction accuracy and generalization. Extensive experimental results on real-world datasets demonstrate that HPAC consistently outperforms established baselines in link-prediction tasks.