A matrix factorization based dynamic granularity recommendation with three-way decisions

Abstract Recommender systems (RSs) are effective technologies and tools used to deal with the problems of information overload, and have been developed rapidly in nearly two decades. In this paper, with the consideration of uncertain and multi-level characteristic of recommendation information (RI), we combine three-way decisions with granular computing, and build a novel dynamic three-way recommendation model to address the limitations of static two-way recommendation. First, we propose a dynamic three-way granularity recommendation (DTWGR), which adopts a top-down dynamic recommendation strategy based on the granular structure of RI. Second, we introduce a matrix factorization framework and further utilize three methods: SVD, SVD++ and NMF, to construct the granular structure of RI. In addition, three extended recommendation algorithms, namely, DTWGR-SVD, DTWGRSVD ++ and DTWGR-NMF, are proposed by combining DTWGR with matrix factorization. The recommendation cost caused by misclassification cost and teaching cost, is carefully investigated in our work. Finally, experimental results of three well-known Movielens data sets validate the effectiveness and reliability of our proposed model.