Parallel context aware recommender system using GPU and JCuda

Abstract

Recommender Systems (RS) are intelligent applications which predict the preferences for users from their current interests. Collaborative Filtering (CF) is a widely used Recommendation technique. On large scale data, collaborative filtering approach is very time consuming and hence parallel processing can be useful for accelerating the task of recommendation. Context Aware Recommender Systems (CARS) help us to generate precise and relevant recommendations by using specific context of the user. This paper presents Parallel Context Aware CF Recommender Systems using JCuda (PCARS) that involves Graphic Processing Unit (GPU). The proposed algorithm works in two phases: offline and online processing. Two kernels are identified and processed for offline processing which reduce the processing time drastically. Online processing involves with the prediction calculation for the items which is not yet seen by target user and can be recommended to them. Offloading the prediction calculation to GPU will help to increase the overall system performance significantly. A prototype of the proposed system is developed using JCuda, CUDA and Java technologies for the restaurant domain. The performance of the proposed system PCARS is compared to the Collaborative Filtering Recommender Systems (CFRS) without parallel processing using contextual pre-filtering and context post-filtering in terms of processing time. Experimental results demonstrated the tremendous speedup of the system over a single-core CPU based system.