End-to-End Learning of Graph Similarity

Constructing and calculating the metrics of graphs comparison precisely can be expensive due to the prohibitively high time complexity, exponential in some cases. Thus building a learning model to approximate the metrics is expected. In this paper, we convert the computation of graphs similarity/distance into a learning problem and propose an end-to-end GCN(Graph Convolutional Network) based model to calculate the GFD(Graphlet Frequency Distribution) distance of graphs. In this way, the trained model predicts the GFD distance of graphs directly rather than constructs a GFD vector by counting graphlets as in traditional methods. A experimental evaluation is conducted to validate the effectiveness of our model in real-world networks scaled from tens of nodes to thousands of nodes. Our trained model takes 480times less time on average compared with the count-based method in the dataset. The 3-top nearest accuracy reaches 74.6% while the 5-top nearest accuracy reaches 85.2% in the test data.