Analysis of Social Networks by Tensor Decomposition

Abstract

The Social Web fosters novel applications targeting a more efficient and satisfying user guidance in modern social networks, e.g., for identifying thematically focused communities, or finding users with similar interests. Large scale and high diversity of users in social networks poses the challenging question of appropriate relevance/authority ranking, for producing fine-grained and rich descriptions of available partners, e.g., to guide the user along most promising groups of interest. Existing methods for graph-based authority ranking lack support for fine-grained latent coherence between user relations and content (i.e., support for edge semantics in graph-based social network models). We present TweetRank, a novel approach for faceted authority ranking in the context of social networks. TweetRank captures the additional latent semantics of social networks by means of statistical methods in order to produce richer descriptions of user relations. We model the social network by a 3-dimensional tensor that enables the seamless representation of arbitrary semantic relations. For the analysis of that model, we apply the PARAFAC decomposition, which can be seen as a multi-modal counterpart to common Web authority ranking with HITS. The result are groupings of users and terms, characterized by authority and navigational (hub) scores with respect to the identified latent topics. Sample experiments with life data of the Twitter community demonstrate the ability of TweetRank to produce richer and more comprehensive contact recommendations than other existing methods for social authority ranking.