| In the real world,static and dynamic graphs are two fundamental types of graphs extensively applied in various fields,such as biomolecular networks and social networks,etc.Consequently,extracting the latent information of these graphs is of high pratical value for solving real-world problems.For instance,in biomolecules,static graphs can represent the structure and function of a molecule,while in social networks,dynamic graphs can reflect the relationships among users at different times.Currently,graph classification task is an important research direction for mining graph information.The main approachs in this field involve extracting features from nodes and edges within graphs and converting them into vector representations of the entire graph for graph classification.However,these methods still face two challenges: 1)existing models ignore the differences of nodes in discriminability for the graph,resulting in ineffectively capture the relationship between the part(node)and the whole(graph);2)existing models neglect the evolutionary patterns of dynamic graphs at different times,leading to the inability to capture the graph evolution process completely.To address these issues,this paper conducts the following research:(1)To tackle the issue that existing methods cannot effectively capture the relationship between the part(node)and the whole(graph)in static graphs,we propose the Multi-Head Routing Graph Capsule Network(MHR-GCAPS).In particular,to capture the partial information of the graph,the model employs the multi-layer graph convolutional network with residual connections to extract node information and combine it into node capsules.To accurately depict the part-whole relationship,we propose the Multi-Head Routing(MHR)mechanism that employs a learnable routing network to capture the relationship between node capsules and graph capsules.Experiments conducted on four biological datasets and three social network datasets show that the model outperforms state-of-theart methods in the graph classification task.(2)To address the inability of existing models to capture the complete evolution process of dynamic graphs,we propose the Generative Adversarial Dynamic Graph Convolutional Network(GADGCN),consisting of the generator and the discriminator.In particular,the generator is a dynamic graph convolutional network composed of the structural encoder and the temporal encoder.In the structural encoder,the graph convolutional network and gated recurrent units are utilized to capture structural and temporal information of nodes.In the temporal encoder,to reduce noise interference in the hidden state of the long short-term memory network and avoid affecting the evolution inference of the entire temporal network,the model uses a shared decoder to model the evolutionary patterns of each hidden state at different times.Furthermore,to prevent instability in the generative module from hindering the model’s effective capture of temporal information,the model further uses the discriminator to optimize the entire evolution process.Experimental results on the USCB and SYN datasets demonstrate that the model enhances the modeling ability for dynamic graph evolution. |
