Parallel collaborative filtering recommendation model based on two-phase similarity

Abstract

Problems such as cold startup, accuracy, and scalability are faced by traditional collaborative filtering recommendation algorithm if the system is expanded continuously. To resolve these issues, we propose a parallel collaborative filtering recommendation model on the basis of two-phase similarity (PCF-TPS) and weighted distance similarity measure (WDSM). In accordance with WDSM, the users’ similarity is calculated and their similarity matrix is obtained. At the same time, the items’ similarity is counted and its similarity matrix is got in line with Tanimoto Coefficient Similarity. For the users’ similarity matrix, their preferences are endowed with weights and in this way their new preferences matrix is received. In addition, the nearest neighbor item is found and a more accurate recommendation to the target user is given on the basis of the items’ similarity matrix and users’ new preferences matrix. Besides, in regard to the parallel computing framework, the parallel implementation of the model is completed. All these experiments are done on MovieLens dataset. The results show that PCF-TPS solves the problem of cold startup and increases the accuracy concerning CF. Compared with PCF-EV, PCF-TPS’s parallel realization can be improved to nearly 125 times on the whole. That is to say, it will be more meaningful to complex model using GPU than a small model. What’s more, PCF-EV’s distributed implementation is much more efficient than PCF-EV’s.