A systematic comparison of community detection algorithms for measuring selective exposure in co-exposure networks

The use of community detection techniques for understanding audience fragmentation and selective exposure to information has received substantial scholarly attention in recent years. However, there exists no systematic comparison, that seeks to identify which of the many community detection algorithms are the best suited for studying these dynamics. In this paper, I address this question by proposing a formal mathematical model for audience co-exposure networks by simulating audience behavior in an artificial media environment. I show how a variety of synthetic audience overlap networks can be generated by tuning specific parameters, that control various aspects of the media environment and individual behavior. I then use a variety of community detection algorithms to characterize the level of audience fragmentation in these synthetic networks and compare their performances for different combinations of the model parameters. I demonstrate how changing the manner in which co-exposure networks are constructed significantly improves the performances of some of these algorithms. Finally, I validate these findings using a novel empirical data-set of large-scale browsing behavior. The contributions of this research are two-fold: first, it shows that two specific algorithms, FastGreedy and Multilevel are the best suited for measuring selective exposure patterns in co-exposure networks. Second, it demonstrates the use of formal modeling for informing analytical choices for better capturing complex social phenomena.