An End-to-End Heterogeneous Network for Graph Similarity Learning

Abstract

At present, the similarity learning methods used in many deep models can only capture the local context relations among samples in small cliques (e.g. pairs and triplets), so it is difficult to construct and leverage their latent global correlations to further boost the learning of discriminant features. To this end, we propose a novel end-to-end heterogeneous network, in which a global correlation inference model is adopted to induce the learning of a feature embedding model. Specifically, we design an intersection graph structure to build local context similarity for each sample. Then, taking the learned feature embeddings as nodes, the connection relationship between all nodes is computed through the global correlation inference model. Meanwhile, by combining two loss functions it can feedback the global connection information to push the discriminant learning of the feature embedding model. Consequently, both parts of the proposed heterogeneous network can be optimized from each other during deep similarity learning. The extensive ablation studies and comparative experimental results demonstrate that each component is effective and the whole network is competitive.