Key Vertex Pairs Extraction for Graph Classification Based on Probability Distribution Learning

Graph classification is a hot topic of machine learning for graph-structured data, and it is also a very potential and valuable research. However, the difficulty of graph classification is challenging and special, which is quite different from the normal classification problems. One of the most difficult points of graph classification is that the numbers of vertex neighbors in graphs are usually variable, which makes the number of weights uncertain and ambiguous. Recent work such like the graph attention network apply the transformer on the graph neural network. However, the learned attentions cannot strictly reveal the importance of each part of graph, which makes the model less explainable. Moreover, for small datasets, it performs less effectively because of the excessive parameters. In order to overcome these difficulties, we propose a lightweight model with an edge weighting function based on the probability distributions of node pair features learned by the Gaussian mixture model. Although the proposed framework is simple, the experimental results shows its effectiveness on small datasets.