Diversity enhancement in community recommendation using tensor decomposition and co-clustering

Abstract

The major aim of Recommender System is to provide appropriate items for user, based on his preferences and intuitively be assessed with accuracy based metrics like precision and recall. Though, diversity of recommended lists is a new emerging debate in RS evaluation. This work tries to improve diversity of community recommendation, using membership as main and tag collection as complementary resource. With exploiting Tensor Decomposition and using Latent Semantic Analysis, communities can be represented in latent topics, based on different modes including member-users and tag-collections. As the main contribution, this work applies diversification on recommended list in different modes, based on intuitive idea that, communities can be differ from different points of view such as membership, or tag collections. Experimental results accomplished on a Flicker dataset show the meaningful improvement in aggregate diversity (for the system) with less accuracy-loss comparing to current methods; moreover it also shows improvements in intra-list diversity (for single user) which is neglected in previous works. As a result, clustering the communities with similar users, or tags, gives the opportunity to diversify the recommended lists to cover more diverse communities with different member users, or different tag content, and this multi-mode diversity lead to better list for user and better coverage for system.