An improved matrix factorization model under multidimensional context situation

Traditional recommender systems have won great success on electronic commerce. However, when the user-item rating record matrix is sparse, traditional recommendation algorithms perform poor, which is known as the cold-start problem. Recently, more and more contextual features have been proven to be valuable information for improving the accuracy of recommendation, and newly formed context-aware recommendation systems (CARS) provide a way to solve the cold-start problem by using some certain features such as users location, mood and social relationship. In order to handle multidimensional context, this paper first extracts relevant contextual information by calculating the information entropy, then divides the contextual information into three categories - user context, item context and interaction context. Finally, we extend the matrix factorization (MF) model to integrate the context information. Experimental results on LDOS-CoMoDa dataset have shown that our approach provides improvement in terms of recommendation accuracy.