Uncovering latent structure in social networks using graph embeddings

Abstract

The data from real-world social networks is huge, heterogeneous, and most of the attributes are unstructured. As a result, extensive investigation is necessary to transform this data into a useful and informative format. Failing to process these unstructured attributes results in a massive loss of potential. Unstructured attributes can offer crucial additional context. Identifying communities and clusters in social networks has been one of the commonly studied problems of graph mining, and is recognized as a challenging necessary task, and many open tasks are still poorly understood. We show that user information from social network platforms such as Instagram can be clustered using similarities based independently on profiles, hashtags, images, and explicit links. These similarity measures are converted to graphs and then embedded in geometric spaces using spectral embedding techniques. Communities and clusters in these geometric spaces correspond to groups of users with similar interests, and such groups can be used for, for example, targeted marketing, content recommendations, and creating a higher level of customer support. We show in this thesis how to get the data, process attributes, and combine structural information existing in the social network. We present a way to link the different types of attributes with structural information to produce the full context of each user. Embedding techniques are employed to represent the subgraphs in a mutually consistent manner while preserving their entirety. Our approach shows that similarity is better represented using attributes and structural information together than by using any one of these independently, or by merging them all. Our approach outperforms all of the conventional community detection algorithms, often by a large margin. The clusters that are produced do contain users with similar interests since topic models perform well on individual clusters.