Joint Learning of Hierarchical Community Structure and Node Representations: An Unsupervised Approach

TL;DR

This paper introduces Mazi, an unsupervised algorithm that jointly learns hierarchical community structures and node representations, iteratively improving both to enhance graph analysis tasks.

Contribution

Mazi is the first method to simultaneously optimize community detection and node representations in an unsupervised, hierarchical manner.

Findings

Mazi outperforms existing methods on synthetic and real-world graphs.

Joint optimization improves community detection and node embedding quality.

Method enhances link prediction and node classification accuracy.

Abstract

Graph representation learning has demonstrated improved performance in tasks such as link prediction and node classification across a range of domains. Research has shown that many natural graphs can be organized in hierarchical communities, leading to approaches that use these communities to improve the quality of node representations. However, these approaches do not take advantage of the learned representations to also improve the quality of the discovered communities and establish an iterative and joint optimization of representation learning and community discovery. In this work, we present Mazi, an algorithm that jointly learns the hierarchical community structure and the node representations of the graph in an unsupervised fashion. To account for the structure in the node representations, Mazi generates node representations at each level of the hierarchy, and utilizes them to influence the node representations of the original graph. Further, the communities at each level are discovered by simultaneously maximizing the modularity metric and minimizing the distance between the representations of a node and its community. Using multi-label node classification and link prediction tasks, we evaluate our method on a variety of synthetic and real-world graphs and demonstrate that Mazi outperforms other hierarchical and non-hierarchical methods.