Common Subgraph Extraction based on Link Prediction

Abstract

Many state-of-the-art methods for graph classification are based on the graph convolution framework and the message-passing mechanism, which tends to use a convolution-like operation to aggregate the features of vertex neighbors and pass the information to nearby vertices. However, recent researches reveal that there usually exists a heavy batch noise of graphs because of the diverse graph structures, which means that not all parts of a graph is useful for classification, most of them contain a huge amount of noise, which will be also aggregated into vertex features when doing graph convolutions and graph poolings. To overcome these difficulties, we propose a generative graph model that focuses on the link prediction problem of a given set of vertex. Through a loop of predicting link, we can construct one or several common subgraphs from the given graph set, which helps the graph classification.