Multi-Perspective Semantic Enhanced Model for Heterogeneous Graph Embedding with Contrastive Learning

Heterogeneous information network embedding aims to learn node representations from a low-dimensional space while preserving the network's complex structural and semantic information. Recently, inspired by the successful application of contrastive learning in computer vision, graph contrastive learning additive embedding of heterogeneous information networks has also entered the vision of researchers. The current means of graph expansion somewhat lead to changes in the underlying semantics. Hence, contrast schemes that can adequately capture rich and lossless semantics are still in the exploration stage. Given this, we propose a new Multi-perspective Semantic-enhanced Contrastive Learning framework for heterogeneous information network embedding (MSCL). Specifically, we design the contrast goals separately from global and local perspectives. First, we capture complementary semantic information for heterogeneous graphs using two sub-structures, network schema and meta-path. Then, considering that different meta-paths also represent different higher-order semantics from each other, we specifically design contrast targets within the meta-path perspective to learn the independence and semantic consistency between them. In addition, we propose a positive sampling strategy to improve the confidence level of the positive sample set in both semantic and structural dimensions. Extensive experiments on three real-world networks show that the proposed approach outperforms state-of-the-art methods.