Multi-Scale Neighbor Contrastive Learning with Attention Mechanism

Graph Contrastive Learning (GCL) can learn effective representations of graphs in a self supervised manner. However, existing manual graph enhancement strategies that randomly perturb the topology of the graph will not be able to maintain task related information intact. To address the insufficient capture of high-order semantic information in graph representation learning, this paper proposes an Attention-based Multi-scale Neighbor Contrastive Learning (AMNCL) framework. First, a multi-head graph attention mechanism is designed to generate augmented views by adaptively aggregating multi-scale neighborhood features. Second, a hierarchical contrastive loss is constructed to explicitly model the semantic relationships between nodes and neighbors at different scales. Finally, learnable weight coefficients are introduced to dynamically balance global and local information.