Cluster analysis on dynamic graphs

Abstract

In the era of big data, massive amount of graph data are generated from various domains like citation networks, biological systems, and social networks, leading to the need of effective analysis techniques. Graph clustering (methods for identifying closely connected groups within datasets) has become increasingly popular. Using neural networks has demonstrated potential for achieving effective clustering results. However, dynamic graphs pose unique challenges compared to static ones. Challenges include handling multiple interactions between nodes, evolving node features and cluster structures over time, and the absence of suitable evaluation metrics.

This paper addresses these challenges by refining the dynamic graph clustering algorithm TGC. Our contributions include introducing the "Intensity Modularity" metric for evaluation, implementing innovative training and sampling techniques to enhance TGC's adaptability, and proposing a method for dynamic determination of cluster numbers. Experimental results validate the effectiveness of our approaches.