Community detection in social networks by spectral embedding of typed graphs

Although there is considerable disagreement about the details, community detection in social networks requires finding groups of nodes that are similar to one another, and different from other groups. The notion of similarity is therefore key. Some techniques use attribute similarity—two nodes are similar when they share similar attribute values; some use structural similarity—two nodes are similar when they are well connected, directly or indirectly. Recent work has tried to use both attribute and structural similarity, but the obvious challenge is how to merge and weight these two qualitatively different types of similarity. We design a community detection technique that not only uses attributes and structure, but separates qualitatively different kinds of attributes and treats similarity different for each. Attributes and structure are then combined into a single graph in a principled way, and a spectral embedding used to place the nodes in a geometry, where conventional clustering algorithms can be applied. We apply our community detection technique to real-world data, the Instagram social network, which we crawl to extract the data of a large set of users. We compute attribute similarity from users’ post content, hashtags, image content, and followership as qualitatively different modes of similarity. Our technique outperforms a range of popular community detection techniques across many metrics, providing evidence that different attribute modalities are important for discovering communities. We also validate our technique by computing the topics associated with each community and showing that these are plausibly coherent. This highlights a potential application of community detection in social networks, finding groups of users with specific interests who could be the targets of focused marketing.