Spectral community detection in sparse networks

Spectral methods based on the eigenvectors of matrices are widely used in the analysis of network data, particularly for community detection and graph partitioning. Standard methods based on the adjacency matrix and related matrices, however, break down for very sparse networks, which includes many networks of practical interest. As a solution to this problem it has been recently proposed that we focus instead on the spectrum of the non-backtracking matrix, an alternative matrix representation of a network that shows better behavior in the sparse limit. Inspired by this suggestion, we here make use of a relaxation method to derive a spectral community detection algorithm that works well even in the sparse regime where other methods break down. Interestingly, however, the matrix at the heart of the method, it turns out, is not exactly the non-backtracking matrix, but a variant of it with a somewhat different definition. We study the behavior of this variant matrix for both artificial and real-world networks and find it to have desirable properties, especially in the common case of networks with broad degree distributions, for which it appears to have a better behaved spectrum and eigenvectors than the original non-backtracking matrix.