Topic-based personalized recommendation for collaborative tagging system

Collaborative tagging has become a very popular way to share, annotate, and discover online resources in Web 2.0. Yet as the number of resources in Collaborative tagging system grows over time, sifting through the large amounts of resources and finding the right resources to recommend to the right user is becoming a challenging problem. In this paper, we investigate a probabilistic generative model for collaborative tagging, explore the implicit semantic connections in the sparse and noisy information space of heterogeneous users and unsupervised tagging. First, a modified Latent Dirichlet Allocation (LDA) model is used to cluster the tags and users simultaneously. The generalization of resource description and user could alleviate the tag noise and data sparseness of recommendation effectively. And then, considering that topic-based recommendation only takes the users' global interest into consideration without the capability of distinguishing users' interest in detail, we combine the global interests with the individual interest and community interest. Experimental results demonstrate the topic-based personalized recommendation method, which integrate both the commonality factor among users and the specialties of individuals, could alleviate data sparsity and provide a more flexible and effective recommendation than previous methods.