Graph Embedding Based on Characteristic of Rooted Subgraph Structure.

Given the problem that currently distributed graph embedding models have not yet been effectively modeled of substructure similarity, biased-graph2vec, a graph embedding model based on structural characteristics of rooted subgraphs is proposed in this paper. This model, based on the distributed representation model of the graph, has modified its original random walk process and converted it to a random walk with weight bias based on structural similarity. The appropriate context is generated for all substructures. Based on preserving the tag features of the nodes and edges in the substructure, the representation of the substructure in the feature space depends more on the structural similarity itself. Biased-graph2vec calculates the graph representations with unsupervised algorithm and could build the model for both graphs and substructures via universal models, leaving complex feature engineering behind and has functional mobility. Meanwhile, this method models similar information among substructures, solving the problem that typical random walk strategies could not capture similarities of substructures with long distance. The experiments of graph classification are carried out on six open benchmark datasets. The comparison among our method, the graph kernel method, and the baseline method without considering the structural similarity of long-distance ions is made. Experiments show that the method this paper proposed has varying degrees inordinately improved the accuracy of classification tasks.