Covariate-Assisted Community Detection on Sparse Networks

TL;DR

This paper introduces a new model and spectral clustering method that effectively detects communities in sparse networks by integrating covariate information, even when covariate labels are inconsistent with community labels.

Contribution

It proposes the DCSBM-NC model and the CA-SCORE algorithm, which together improve community detection in sparse networks using covariates, handling high-dimensionality and label inconsistency.

Findings

CA-SCORE accurately recovers communities in synthetic data

Method performs well on real datasets

Effective in networks with dense connections and matching covariate labels

Abstract

Community detection is an important problem when processing network data. Traditionally, this is done by exploiting the connections between nodes, but connections can be too sparse to detect communities in many real datasets. Node covariates can be used to assist community detection; see Binkiewicz et al. (2017); Weng and Feng (2022); Yan and Sarkar (2021); Yang et al. (2013). However, how to combine covariates with network connections is challenging, because covariates may be high-dimensional and inconsistent with community labels. To study the relationship between covariates and communities, we propose the degree corrected stochastic block model with node covariates (DCSBM-NC). It allows degree heterogeneity among communities and inconsistent labels between communities and covariates. Based on DCSBM-NC, we design the adjusted neighbor-covariate (ANC) data matrix, which leverages covariate information to assist community detection. We then propose the covariate-assisted spectral clustering on ratios of singular vectors (CA-SCORE) method on the ANC matrix. We prove that CA-SCORE successfully recovers community labels when 1) the network is relatively dense; 2) the covariate class labels match the community labels; 3) the data is a mixture of 1) and 2). CA-SCORE has good performance on synthetic and real datasets. The algorithm is implemented in the R(R Core Team (2021)) package CASCORE.