Exploring generative models of tripartite graphs for recommendation in social media

As social media sites grow in popularity, tagging has naturally emerged as a method of searching, categorizing and filtering online information, especially multimedia content. The unrestricted vocabulary users choose from to annotate content however, has often lead to an explosion of the size of space in which search is performed. This paper is concerned with investigating generative models of social annotations, and testing their efficiency with respect to two information consumption oriented tasks. One task considers recommending new tags (similarly new resources) for new, previously unknown users. We use perplexity as a standard measure for estimating the generalization performance of a probabilistic model. The second task is aimed at recommending new users to connect with. In this task, we examine which users' activity is most discriminative in predicting social ties: annotation (i.e. tags), resource usage (i.e. artists), or collective annotation of resources altogether. For the second task, we propose a framework to integrate the modeling of social annotations with network proximity. The proposed approach consists of two steps: (1) discovering salient topics that characterize users, resources and annotations; and (2) enhancing the recommendation power of such models by incorporating social clues from the immediate neighborhood of users. In particular, we propose four classification schemes for social link recommendation, which we evaluate on a real--world dataset from Last.fm. Our results demonstrate significant improvements over traditional approaches.