Contrastive Graph Similarity Networks

Graph similarity learning is a significant and fundamental issue in the theory analysis of graphs, which has been applied variety fields, including object tracking, recommender systems, search, etc. Recent methods for graph that utilize deep typically share two deficiencies: (1) they leverage neural networks as backbones representations but have not well captured complex information inside data, (2) employ cross-graph attention mechanism learning, computationally expensive. Taking these limitations into consideration, method devised this study, namely, C ontrastive G raph Sim ilarity Network (CGSim). To enhance CGSim makes use complementary input graphs captures pairwise relations contrastive framework. By developing dual module with node-graph matching graph-graph mechanism, our significantly reduces quadratic time complexity interaction modeling to linear complexity. Jointly an end-to-end framework, representation embedding well-designed can be beneficial one another. A comprehensive series experiments indicate outperforms state-of-the-art baselines on six datasets computational cost, demonstrates model’s superiority over other baselines.