Probabilistic Generative Models of the Social Annotation Process

With the growth in the past few years of social tagging services like Delicious and CiteULike, there is growing interest in modeling and mining these social systems for deriving implicit social collective intelligence. In this paper, we propose and explore two probabilistic generative models of the social annotation (or tagging) process with an emphasis on user participation. These models leverage the inherent social communities implicit in these tagging services. We compare the proposed models to two prominent probabilistic topic models (Latent Dirichlet Allocation and Pachinko Allocation) via an experimental study of the popular Delicious tagging service. We find that the proposed community-based annotation models identify more coherent implicit structures than the alternatives and are better suited to handle unseen social annotation data.