The Research Of Distributed Collaborative Filtering For Cold-Start Recommendations

As an important part of Recommender System (RS), Collaborative Filtering (CF)has been widely used. CF scheme searches similar items or users from the item-usermatrix and predicts the preference of users to assist recommender systems in recommendingitems that they are likely to be of interest to users.Various CF schemes have been proposed, such as memory-based schemes, modelbasedschemes and hybrid schemes. These schemes have greatly improved the performanceof RSs. However, many traditional CF schemes suffer from sparsity and scalabilityproblems. Sparsity problem means that the item-user matrices are extremelysparse and the performances and precisions of CF schemes are low. Scalability problemis the fact that the computational complexity is too high for CF schemes to deal withlarge item-user matrices with hundreds of millions of items or users. Moreover, withoutenough information, most CF schemes are seriously limited by a cold-start problemwhich refers to a situation that RSs are incapable of drawing recommendations for newitems, new users or both.In this paper, we propose a distributed collaborative filtering scheme for cold-startrecommendations. It identifies and filters insignificant ratings by introducing ratingconfidence level, which helps RSs to deal with large-scale item-user matrices. To constructa dense area, the scheme identifies popular items and frequent raters and swapsthem to the most upper left of the item-user matrix. By using matrix singular valuedecomposition, the dimensionality of the item-user matrix is substantially reduced.To overcome data sparsity, the scheme co-clusters items and users, and smoothes ratingswithin every user cluster. To make recommendations, the scheme predicts userpreference by fusing recommendations from item and user clusters. It calculates simi- larities by using mutual information so that the nonlinear correlations of items or usersare evaluated and the precision of the scheme is also increased. To further improvethe performance, the CF system is implemented on MapReduce distributed computingframework. Experiments on MovieLens and Netflix dataset show that the CF schemesolves the cold-start problem regarding recommendation accuracy and scalability andfits current RSs well.