Graph Representation Learning In A Contrastive Framework For Community Detection

摘要

Graph structured data has become very popular and useful recently. Many areas in science and technology are using graphs for modeling the phenomena they are dealing with (e.g., computer science, computational economics, biology, …). Since the volume of data and its velocity of generation is increasing every day, using machine learning methods for analyzing this data has become necessary. For this purpose, we need to find a representation for our graph structured data that preserves topological information of the graph alongside the feature information of its nodes. Another challenge in incorporating machine learning methods as a graph data analyzer is to provide enough amount of labeled data for the model which may be hard to do in real-world applications. In this paper we present a graph neural network-based model for learning node representations that can be used efficiently in machine learning methods. The model learns representations in an unsupervised contrastive framework so that there is no need for labels to be present. Also, we test our model by measuring its performance in the task of community detection of graphs. Performance comparing on two citation graphs shows that our model has a better ability to learn representations that have a higher accuracy for community detection than other models in the field.