Evaluating Community Detection Algorithms for Multilayer Networks: Effectiveness of Link Weights and Link Direction

Analyzing the structures of multilayer networks (MLNs) has been a hot research topic in network science. Community detection algorithms are important tools for analyzing MLNs. In the literature, several community detection algorithms for MLNs have been proposed. Moreover, there are several options for the graph representation of an MLN: for example, directed or undirected, weighted or unweighted, and using information from all or only some layers. Although these options may affect the results of community detection in MLNs, representations that are effective for community detection have not yet been clarified. In this paper, we experimentally evaluate the effectiveness of three types of community detection algorithms for MLNs and examine how the graph representation of an MLN affects the results of these algorithms. Our main findings are as follows: (1) The flattening approach is particularly effective, whereas the layer-by-layer approach is not applicable to detecting communities in MLNs of Twitter users. (2) Using a directed graph for each layer of an MLN increases the accuracy of community detection. (3) The Leiden method, which is a community detection algorithm for single-layer networks, achieves comparable accuracy with the community detection algorithms for MLNs, which suggests that there exists room for improvement in multilayer community detection algorithms for effectively utilizing the multilayer structures of MLNs.