Higher-order Structural Embedding And Composite Pooling Integrated Graph Classification Model

Graph classification is an important research direction in the field of graph mining and is applied in many practical areas.Common data in the real world,such as social networks,urban traffic networks,chemical molecular networks,etc.,can often be represented as graph structures.Classification tasks based on graph structure data,such as determining the physicochemical characteristics of chemical molecular structures and using protein molecular structures for new drug discovery,are therefore of great importance to achieve accurate graph classification using existing neural network models.In recent years,graph neural networks have achieved great success in node classification and link prediction.In addition to nodes containing a large amount of feature information,higher-order structures represented by sets of two or more nodes in a diagram are also rich in semantic information,such as the functional groups in a compound molecule that are important for analysing the properties of the whole molecule.However,when using graph neural networks for graph classification,the general approach is to generate embedding representations for all nodes in the graph and then globally aggregate these node embeddings as features of the graph.This approach has two limitations:first,it uses only the information of vertices and edges in the graph for classification,lacking attention to higher-order graph structure information,and second,the model lacks hierarchical structure.The naturally composed network structure itself is formed from individual nodes through interconnections,which contains a large amount of structural semantics,and learning the graph representation in a hierarchical manner is important to capture the local structure present in the graph.In order to develop effective graph classification models,it is necessary to make full use of the rich information inherent in graph structures and the feature information contained in graph nodes and edges.To address the above issues,we propose a graph classification model based on higher-order structure embedding and a graph classification model based on composite pooling.Firstly,we use an incremental processing process by aggregating higher-order structure information order by order,and use a new compound pooling mechanism in the hierarchical model to give a subgraph importance score based on node features and higher-order structure information synthetically for the pooled subgraph nodes,preserving the semantic information on the multi-order structure,and finally obtaining a more complete final feature representation.The innovations of this thesis are as follows:(1)Selecting key nodes in the graph to incrementally embed higher-order structure information,which makes full use of the graph structure information and avoids the redundancy caused by computing against a large number of nodes.The key to utilizing higher-order structures is that they pass messages directly between subgraphs rather than between individual nodes,and this higher-order form of message passing captures structural information that is not visible at the node level.(2)Considering both subgraph feature information and graph topology information in the pooling process,for feature generation after selecting pooled subgraphs,feature fusion is used to ensure that the feature representation of pooled subgraphs contains sufficient valid information from the graph.(3)Further,a new graph neural network model fusing higher-order structural embedding and composite pooling is proposed to fuse the two to deal with the graph classification problem,and the model effect is validated on a commonly used benchmark dataset,and the effects of higher-order structural embedding and pooling are analyzed and verified by ablation experiments,respectively.Experimental results on publicly available graph classification datasets show that the feature augmentation algorithm incorporating composite pooling has significant advantages over existing methods in terms of classification effects and can better perform in real classification tasks.In the next research work,we can conduct research on realistic graph data in different domains to promote the implementation of graph classification in various professional fields.