Multi-Scale Contrastive Learning Based Heterogeneous Graph Embedding

Recently, heterogeneous graph contrastive learning has been widely explored as a self-supervised solution for heterogeneous information network embedding. However, based on contrastive learning, most existing graph expansion proposals fail to extract heterogeneous information adequately, and the simple contrastive objective neglects the modelling of the consistency factor among views. In addition, the single criterion of dividing samples makes the problem of sampling bias in contrastive learning obvious. In response to the issues to be solved, we present a multi-scale contrastive learning method (MCLHE) applied to heterogeneous graph embedding. Concretely, we initially set two scales of contrast objectives for cross-schema and cross-view, capturing the consistency factor between complementary information sources. Secondly, we design node-level contrast and graph-level contrast within the contrast objectives, which are combined to convey different granularity of information. Finally, we take into account both structural and feature embedding similarity measures in the sample partitioning process to enhance sampling hardness. Extensive experiments on several public datasets indicate that MCLHE obtains better performance than existing methods.