Differentiable Community Detection with Graph Neural Networks and Stochastic Block Models

William Arliss, W. Graham Mueller

Keywords: Stochastic Block Models, Graph Neural Networks, Unsupervised Learning, Community Detection

TL;DR: This paper proposes a set of loss functions adapted from Stochastic Block Model likelihood functions to train Graph Neural Networks for the task of unsupervised community detection.

Abstract: We propose a set of loss functions adapted from Stochastic Block Model (SBM) likelihood functions to train Graph Neural Networks (GNNs) for the task of unsupervised community detection. Identifying latent community structures is a prominent challenge for many graph applications. SBMs are classical models that describe the generating process of random graphs and are commonly used to infer community structure. The likelihood functions associated with SBMs are well-defined, differentiable, and measure the quality of inferred community partitions; this makes them particularly useful for unsupervised learning with GNNs. Our proposed loss functions are independent of any specific GNN architecture and demonstrate competitive or improved community detection performance against several alternatives. Evaluation is carried out with multiple architectures, offering a thorough empirical analysis of the state of community detection with GNNs.

Submission Number: 5