Research On Hierarchical Architecture Graph Classification Based On Deep Learning

As a type of particular data structure,graphs can serve as excellent representations for a vast number of sophisticated networks in real-world scenarios.Graph classification also has emerged as a top research currently as its applications with a widespread field of applications in actual cases,such as protease detection,traffic network prediction,and drug activity analysis.With the advancement of the technology of Graph Neural Networks,the graph classification method based on deep learning has advanced significantly,which extracts the feature information and reduces the dimension of the graph by employing the graph convolution layer and the graph pooling layer,respectively,followed by a readout layer to get the graph representation.As a result,the classification performance is significantly upgraded.Nevertheless,there still exist some issues in such methods,which include not making proper usage of the structural information of the graph in the graph convolution layer,discarding nodes in the graph pooling layer resulting in missing information,as well as existing information redundancy in the graph readout layer.In consequence,targeting the aforesaid issues,the work done in this study is as follows:(1)Graph classification approach using a hierarchical architecture based on graph pooling and adaptive readout.For the sake of more thoroughly integrating node information during pooling operations as well as diminishing information redundancy during readout,a hierarchical architectural graph classification method based on graph pooling and adaptive readout is suggested in this thesis.Firstly,the information carried by nodes is enhanced by paying attention to the first-order neighbor node information as well as the second-order neighbor node information with respect to the node score calculation.Then,a system for the dynamic fusion of node scores is designed so that it is possible for nodes to focus on more vital information scores in obtaining the final score and eliminates the arbitrary subjectivity of artificial weight assignment.As for the coarse graphs generated by pooling,this research incorporates the attention mechanism to improve the readout layer and designs an adaptive readout layer,whereby the graph representations of each coarse graph can be fused through the attention mechanism so as to better utilize the information of these graph representations as well as prevent information duplication.Some experiments are done on graph classification datasets,and the experimental outcomes prove the validity of the model.(2)Graph classification approach using a hierarchical architecture based on the high-level structure and result fusion.For the sake of more comprehensively considering the structural information of the graph and preserving better information integrity,a hierarchical architectural graph classification method based on high-level structure and result fusion(HLRF)is proposed in this thesis.To begin with,when selecting the feature information of nodes,the messages of all neighboring high-level nodes of the current node are aggregated on that node,which allows the model to pick up more graph information.Then,this thesis variants the traditional hierarchical architecture and designs a result fusion module.In this module,it allows each readout layer to be immediately followed by a classifier to make predictions for each graph representation in advance,and then integrates these predict results to yield the final graph classification results.Such initiatives mentioned above enable HLRF to make better use of the graph structure,at the same time,applying the result fusion module permits the model to maintain more accurate information about the nodes.With several experiments,it is demonstrated that the HLRF model shows sound classification performance.