Deep Graph Discriminant Learning

Graph is a data structure that can describe the complex relationships between things.Graph discriminant learning aims to use the structural information of graphs and feature of nodes to obtain more robust and discriminant graph feature representation,so as to better support various downstream tasks,such as social networks,biomedicine,recommendation systems and various computer vision tasks.Although the development of graph convolutional neural network has made great progress in graph discriminant learning,there are still many important problems that have not been effectively solved.For example,the structural learning of graphs can only rely on fixed and explicit structure information,which leads to the structure representation of graphs being too shallow.There is still a lack of highly discriminative learning methods to measure the similarity between homogeneous and heterogeneous graphs.To solve the above problems,this paper proposes solutions from the perspectives of context awareness,Wasserstein measurement and coupled dictionary,and achieves excellent results on the datasets corresponding to multiple tasks.The main innovations of this paper include the following three aspects:· Aiming at the problem of robust representation of graph structure,a graph representation learning method based on context awareness is proposed.By modeling the context information of the current data,the specific implicit graph structural information is learned,and fused with the fixed explicit graph structure rules to realize the context-aware graph representation learning.In addition,the method also considers that the context information has different importance to the current data.So the variational inference method is introduced to model the distribution of context information,and learns the adaptive weight for them.Compared with the existing methods,our method achieves superior performance in multi-label associated image classification tasks.· Aiming at the problem of similarity discrimination between homogeneous graphs,a graph discriminant learning method based on Wasserstein measure is proposed.Since the topological structure of graphs is very important in discriminant learning of graphs,Wasserstein distance is introduced to measure the topological similarity between two graphs.At the same time,in order to enhance the discriminant learning ability of graphs,a dynamic graph dictionary is learned in Wasserstein space and constraints are added to make dictionary keys compact within classes and scattered between classes.This method takes the dictionary as a bridge,and uses Wasserstein distance to measure the topological correlation between the input graph and keys of the dictionary graph.Then,we can ob-tain the concise and high discriminant features of the input graph by projecting the input graph into the dictionary space.According to the obtained feature of the input graph,the similarity between graphs is measured.Compared with the existing methods,our method achieves excellent performance in the task of graph classification.· Aiming at the application of similarity discrimination between heterogeneous graphs,a discriminant learning method of heterogeneous graphs based on coupled dictionary is proposed.Considering that the intrinsic structure of graphs is robust in discriminant learning of heterogeneous graphs,this method constructs a corresponding graph dictionary for each mode and constrains the one-to-one correspondence of graph keys between different modes.Thus a coupled dictionary is formed.Based on the coupled dictionary,different heterogeneous graphs are mapped to the corresponding dictionary spaces by Wasserstein metric,so that they could be encoded into concise vectors.The discrimination between heterogeneous graphs is carried out according to the vectors.At the same time,the method also constructs a novel loss function acting on the coupled dictionary,which makes the coupled key pairs more compact and the uncoupled key pairs more dispersed.Compared with the current advanced methods,the effectiveness of the proposed method has been validated in the application of cross-modal heterogeneous retrieval.