IMCSN: An improved neighborhood aggregation interaction strategy for multi-scale contrastive Siamese networks

Haoyu Zhang, Daoli Wang, Wangshu Zhao, Zitong Lu, Xuchu Jiang

Abstract: Highlights•In this study, uniform directed random noise is added to the learned embedded representations to obtain data enhancement at the representation level while constructing the dyadic subgraphs in parallel and thus preserving the graph structural integrity. Since the granularity of the added noise is controllable, the method can construct embedded representations that are different from the original representations while retaining the learnable information in the original representations, increasing the diversity and robustness of the representations;.•A triple-twin network is used to process the input graphs of the local and global views separately, with GIN as the encoder, and a cross-network, cross-view comparative learning goal is adopted to maximize the consistency of node representations between different network views representation consistency, thus improving the quality and generalization ability of node representations;.•A self-supervised loss function based on graph reconstruction is introduced in the loss function to further optimize the graph representation by using the structural similarity between the original graph and the reconstructed graph, which can make the global view retain more structural information of the original graph, thus improving the complementarity between the global and local views.