UCoDe: Unified Community Detection with Graph Convolutional Networks

Research output: Working paper/Preprint Preprint

62 Downloads (Pure)

Abstract

Community detection is the unsupervised task of finding groups of nodes in a graph based on mutual similarity. Existing approaches for community detection either partition the graph in disjoint, non-overlapping, communities, or return overlapping communities. Currently, no method satisfactorily detects both overlapping and non-overlapping communities. We propose UCoDe, a unified method for unsupervised community detection in attributed graphs. It leverages recent developments in Graph Neural Networks (GNNs) for representation learning. So far, GNN methods for community detection provide competitive results in either overlapping or non-overlapping community detection tasks, but have had little success in both. UCoDe overcomes these issues by introducing a new loss that captures node similarity on a macro-scale. We provide theoretical justification for our approach's validity in the task of community detection and show that it can be applied in both the overlapping and non-overlapping settings. As our experiments demonstrate on several real benchmark graphs, UCoDe consistently provides high quality results in both overlapping and non-overlapping settings in an easy to apply fashion.
Original languageUndefined/Unknown
Publication statusPublished - 29 Dec 2021

Cite this