Heterogeneous Link Prediction via Mutual Information Maximization Between Node Pairs

Yifan Lu, Zehao Liu, Mengzhou Gao, Pengfei Jiao

Abstract: Heterogeneous graphs (HGs), possessing various node and edge types, are essential in capturing complex relationships in networks. Link prediction on heterogeneous graphs has wide applications in real-world. Although existing methods for learning representations of HGs have made substantial progress in link prediction tasks, they primarily focus on the heterogeneous attributes of nodes when capturing the heterogeneity of heterogeneous graphs, therefore, it performs poorly in maintaining pairwise relationships in HG. To address this limitation, we propose a simple yet effective model for link prediction on HGs via Mutual Information Maximization between Node Pairs (MIMNP). We use an Multi-Layer Perceptron as a node encoder to learn node embeddings and maximizes the mutual information between node pairs. Our model effectively preserves the pairwise relationships between nodes, resulting in enhanced link prediction performance. Extensive experiments conducted on three real-world datasets consistently demonstrate that MIMNP outperforms state-of-the-art baselines in link prediction.