A Topology-Based Approach to Pattern Recognition on Graph-Structured Data

摘要

With the ever-increasing scale and complexity of data, the form of graphs will gradually become the mainstream of data storage and display. As a result, the techniques of classification and pattern recognition on graph-structured data have become more and more important and can be applied to a broader range of domains. Based on the recently introduced approach which learned the representations of graphs using convolutional networks, we proposed an improved model, which is called T-PSCN, to classify arbitrary graphs with attributed nodes and edges more effectively. In our model, we integrate the edge attributes as node attributes and exploit topological information at both node and graph level for each graph to enrich the graph representations during the training, which has a significant improvement in accuracy. Results demonstrated that T-PSCN is highly competitive with other graph classification approaches.