Deep Learning Methods On Heterogeneous Graphs

As a relational data structure,graphs which connect data nodes in the form of edges,can effectively solve the problem of modeling large amounts of complex data in real life.Graphs have important applications in social networks,such as biomedicine and knowledge graphs.However,most previous work neglect to distinguish the heterogeneity of objects in the graph and their relationships,which results in a non-negligible loss of information.Heterogeneous graphs are able to model and represent these data,so that we can better fuse not only different types of objects and their interactions,but also multiple modalities.Inspired by deep learning methods,graph convolutional networks try to solve the problem of processing the graph data with complex topological structures.But there are still challenges in directly applying deep learning methods to heterogeneous graphs.This thesis introduces deep learning to obtain a more robust embedding representation and tries to solve the problem of heterogeneous intergraph measurement.This thesis mainly includes the following two aspects:1.To improve the discriminative representations of heterogeneous graphs,this thesis propose a dual-attention graph convolutional network model.By embedding the dual-attention mechanism in the graph convolutional network,short-term and long-term semantic dependencies are introduced from the node level and the semantic level,respectively.Specifically,on the one hand,this thesis proposes the connection attention module on neighboring nodes for aggregating the node-level information.On the other hand,this thesis proposes the hop-attention module on different scale receptive fields in the process of message diffusion,so as to aggregate information of different semantic levels.Thus the proposed model can learn more about the distribution of context.Compared with the classical learning methods,our method achieves excellent performance in semi-supervised node classification and cross-heterogeneous graph retrieval tasks.2.To address the measurement problem between heterogeneous graphs,this thesis proposes a cross heterogeneous graph matching model with the graph Wasserstein correlation analysis,and introduces the graph Wasserstein distance as a measure between two graphs at the first time.In particular,spectral graph filtering is introduced to encode graph signals,which are then embedded as probability distributions in a Wasserstein space,called graph Wasserstein metric learning.The combination of graph signal filtering and graph distance measurement not only captures the similarity of graph signal distribution for better reflecting the topological structures of different heterogeneous graphs,but also retains the transitivity of the embedded space.The theoretical analysis shows that such a seamless integration of graph signal filtering together with metric learning can result in a surprise consistency on both learning processes,and the solution of the proposed model can be derived as a classic generalized eigenvalue decomposition problem.In the experiment,a number of heterogeneous graphs are constructed and cross-compared.The experimental results show that the proposed method achieves the excellent performance in the cross-heterogeneous graph retrieval task.