Multi-Granularity Graph Contrastive Learning Framework via Granular-Ball on Heterogeneous Graphs.
Graph Contrastive Learning (GCL) is a popular self-supervised learning (SSL) technique. However, mainstream GCL methods usually favor single fine-grained random augmentation schemes, which will destroy the structural integrity of the graph, and they largely ignore the topology of the graph structure, that is, multi-granularity characteristics. Graphs are typically composed of homogeneous regions with varying granularities, where nodes within a region exhibit strong homogeneous properties. However, most of the real graphs are heterogeneous, and in the local regions of heterogeneous graphs, interconnected nodes may have similar semantic information, even if they do not belong to the same class. To this end, we propose a new multi-granularity graph contrastive learning framework via granular-ball (GBGCL) to explore the potential on heterogeneous graphs. Specifically, we develop an adaptive granular-ball augmentation strategy that identifies multi-granularity homogeneous regions in the heterogeneous graph, and treat nodes in the same granular-ball as positive pairs, while nodes in different granular-balls as negative pairs. In addition, we integrate the feature information of the nodes in original graph and semantic information of similar nodes in the feature space. Node representations are obtained through joint optimization of losses. Experiments on heterogeneous graphs demonstrate the unique advantages of our framework.