Approximate graph distance with imagisation

Abstract

Graph similarity can contribute to the solutions of a wide variety of real-life problems. Effective graph similarity measures, therefore, are in high demand in areas such as communication network, biology, medicine, finance, etc. Existing similarity methods present key shortcomings including high run-time computational cost and/or strong dependence on feature selection and feature engineering. Many of existing methods also suffer from ambiguity in the interpretation of resultant measurement. In this paper, we propose a fast approximation of graph similarity grounded on convolutional neural network based image embedding. Graph similarity (distance) is, therefore, translated into the quantitative comparison of the corresponding images faithfully encoding structural information of the graphs. The proposed method is validated with purposely built test data. In addition, we have also carried out a preliminary evaluation, that has demonstrated highly promising results: confirming the viability of proposed "imagisation" based graph distance measure.