Community detection in node-attributed social networks: How structure-attributes correlation affects clustering quality

Abstract

The majority of parametric community detection (CD) methods working with node-attributed social networks (ASNs) focus on proposing new techniques and rarely pay much attention on the general analysis how ASN properties affect the corresponding CD quality. However, the latter mostly determines the applicability of a CD method in practice. To fulfil the gap, we investigate CD quality dynamics for ASNs with different structure-attributes correlation. The structure-attributes fusion model under consideration is a weight-based one that interpolates between the so-called fixed and non-fixed topology cases and generalizes a wide class of known weight-based models. Within the model, we first theoretically study the influence of correlation on CD quality and secondly illustrate our conclusions on specially constructed synthetic ASNs. Further, we test our conclusions on original and modified real-world ASNs. Our calculations indicate that the presence of correlation noticeably affects CD quality and that the simultaneous usage of network structure and attributes is not always reasonable within the weight-based fusion model under consideration. This makes the common suggestion that "adding attributes to structure leads to better CD results" questionable in certain cases.