Research On Graph Classification Algorithm With Graph Convolutional Networks Based On Subgraph Feature Distribution

Graph classification as a kind of complex classification task that focuses on the data with irregular topological structures widely exists in numerous interdisciplinary research fields such as social network analysis,machine vision,brain science,and Biochemistry.Since the nonlinear structures contained in the data are difficult to be expressed and calculated in the Euclidean space directly,this problem cannot be solved easily by traditional machine learning methods.Recently,graph convolution networks(GCNs)provide an effective way to generate meaningful embeddings for such graph-structured data,so that they have achieved remarkable achievements in a series of tasks such as node classification and link prediction.However,due to the lack of node order and the variability of structural scale,the existing methods still have some imperfections when being applied to graph classification tasks,e.g.,over-compression of structural information,misalignment of structural semantics,and underutilization for the task characteristics.In this dissertation,we attempt to introduce subgraph feature distribution into graph convolutional networks,and to utilizes attention mechanism and ?₀ regularization techniques to address these problems.The main contributions of this dissertation can be included as following two parts:(1)This dissertation first proposes a graph classification model by leveraging subgraph feature distribution,namely SFD.Firstly,the subgraph feature distribution information is introduced into the graph-level representation by constructing a non-collapse-type read-out strategy.This mechanism can effectively avoid the problem of losing the local structural information caused by overcompression? Secondly,SFD employs a batch strategy to discretize the embeddings of local structures by their structural semantics,which can provide interactive information to approximately align the local-structural embeddings for varying graphs.According to the Wasserstein metric,the aligned structural information is beneficial to capture the similarity among graphs.Lastly,a series of experimental results on a range of graph datasets show that SFD can effectively learn both the local and global-level structural information of graph data,and outperform lots of state-of-the-art graph classification algorithms.(2)Based on SFD,this dissertation further proposes a graph classification network with enhanced subgraph feature learning,namely ESL.Firstly,in consideration of different scales of subgraph structures,adaptive multihop graph convolution is designed.This operation leverages the attention mechanism to learn the structural features of multi-scale subgraphs,which improves the ability of SFD in learning the key local-structural information.Secondly,in consideration of the sparsity of subgraph structures and the character-istics of graph classification tasks,a densesparse parallel graph convolution architecture is constructed.This architecture enhances subgraph features by leveraging ?₀ regularization technique,where more important sparse structures are included in the subgraph feature distribution statistics in SFD.Thus,it is more conducive to the global structural similarity measure based on the subgraph feature distribution.Lastly,A series of experimental results on a range of graph classification datasets show that ESL can significantly improve the ability of SFD in dealing with graph classification problems compared with GCN,GIN,and other graph classification algorithms.