Multi - Layer Community Detection

Abstract

Communities are common structures in social networks. These structures, typically, are formed by different attributes and consequently have different textures in the network. Standard Community Detection (CD) methods detect and extract them to some degree: they often form a node partition clearly related to a dominant node attribute. Such methods are unable to detect the whole variety of communities in the network. We study CD on multi-attributed affiliation networks that are networks with nodes decorated by a number of attributes and edges forming due to their similarity. The networks are represented as a composition of single node attribute networks called one-layer networks and yield node partitions into the attribute clusters. We believe that these partitions can be detected by standard CD algorithms applied to a network accumulated both structural information and node attributes. We propose an iterative method called Multi-Layer Community Detection Algorithm (MLCDA) including two stages a synthesis phase of utilizing available network data and a decomposition phase in which communities are extracted layer by layer. The synthesis includes the conversion of an original network into a weighted one based on assumptions about the network model, construction of an association network accumulated the node attributes, and synthesis of the networks in an accumulated network. In the decomposition phase, CD is conducted on the accumulated network; for the obtained partition an underlying node attribute is determined; an approximation network for the corresponding one layer network is constructed and extracted from the accumulated network; these steps are repeatedly repeated.