Research On Dynamic Graph Representation Learning

In recent years,research on graph data has developed rapidly.Many complex systems in the real world can be represented by graph data structures,such as biological networks,social networks,urban transportation networks,etc.Because these graph data commonly store a lot of valuable information.For example,in social networks,users are divided into different social groups by analyzing their interests and hobbies for correct advertising,user search,and other functions.Therefore,many scholars have researched graph representation learning to explore important information in graph data.With the rapid development of deep learning,Graph Convolutional Network(GCN)as a powerful tool for processing irregular structure data has achieved satisfactory results in node classification,link prediction,and other tasks.However,recent studies have shown that GCN is highly sensitive to the quality of graph structure when it uses a fixed graph structure to guide convolution operations.The quality of graph structure directly affects the effect of subsequent tasks of GCN models.In cases where graph structures are not readily available,many previous works have manually constructed graph structures(such as k NN graph)from input data.However,building similarity between nodes based on straightforward distance measurement is susceptible to noise and outliers,which can lead to poor quality of the graph structure and affect subsequent tasks.In addition,the construction of diagrams is independent of downstream tasks.Fixed diagrams are applied to all layers of the network,and can not capture the underlying structure of node characteristics in different layers very well,resulting in poor model generation.Based on graph learning theory,low rank learning,sparse learning,and other methods,two improved strategies are proposed to solve the problems of graph convolutional network model in graph structure.1)A graph-reconstructed convolutional network(LRGCN)based on Laplace rank constraint is presented.The classification accuracy of graph convolutional network is low when the graph structure obtained from domain knowledge is damaged or even unavailable.LRGCN combines hypergraph and Laplacian rank constraint learning to get a high-quality graph input to GCN for more accurate classification results.First,a high-order relationship between data is established through hypergraph learning.Then,on the basis of the hypergraph,a new high-quality graph structure with c connected components(where c is the number of classes)is learned with the help of Laplace rank constraint of the graph.Finally,the learned graph structure is input the GCN to guide the subsequent node classification tasks.2)A robust dynamic graph learning convolutional network(RGLCN)is presented.RGLCN is an end-to-end network model.Specifically,this paper presents a robust dynamic graph learning loss function based on smooth vacation,sparse constraint,and strong connectivity constraint.Then,graph learning and graph convolution are integrated into a unified framework.Based on downstream tasks,learn the optimal graph structure dynamically.In this paper,GCN is very susceptible to the quality of graph structure and affects the downstream tasks.Two strategies are proposed to improve the node classification performance of the model.The experimental results show that the proposed method is superior to all the comparison algorithms.In the future,considering the introduction of the multi-view method into the graph convolutional network,we can view the same thing from multiple different perspectives,and combine the ability to extract complex features from the graph convolution.Fusing information from multiple perspectives can considerably enhance the model learning ability.