Acceleration Technology Research Of GPU-based And Distributed CPU-based Collaborative Filtering Recommendation Algorithm

In this era of information explosion, in order to meet the more accurate and convenient needs of individual users, personalized recommendation system emerges and attracts more attention. Among existing applicable technique, collaborative filtering is known as the most widely deployed solution. Due to the fast growing of data size, if traditional collaborative filtering algorithms also run in the stand-alone environment, the calculation process of the stand-alone mode recommendation algorithm will be difficult to meet the immediate needs. According to the scalability of the collaborative filtering recommendation algorithm in the large scale data, how to speed up the running time has become an important research direction. This paper will focus on study of accelerating GPU and Distributed CPU-based collaborative filtering recommendation algorithm, the main research results are as follows:1. CPU-based two types of classical algorithm of collaborative filteringrecommendation have been implemented.In the stand-alone mode, using the MATLAB programming technology to implement CPU-based two types of classical algorithm of collaborative filtering recommendation-- User-Based CF and Item-Based CF. The relevant experimental data will be used as a performance baseline comparison of future accelerated algorithm; at the same time, the algorithm analysis and experimental results verify and make sure the performance bottleneck in the collaborative filtering, which is an important object of study of parallel processing, is truly "targeted".2. GPU-based two types of classical algorithm of collaborative filteringrecommendation have been implemented.Based on GPU calculation model research that high computing density part migrated to the GPU in the process of CF. This paper presents a solving thought to achieve accelerated collaborative filtering recommendation algorithm on the GPU: under the large data, put the calculation of the large time cost and high time complexity onto the GPU, and remain the relative small part serially running on the CPU, and accelerate the whole effect in real time. The experiment result shows that the whole accelerated effect depends on the key steps’ accelerated effect and the version of GPU has a significant advantage in the performance.3. Designing a method of GPU-based implementing two types of classicalalgorithm of collaborative filtering recommendation.Because of using MDCS and PCT to build a distributed computing platform is extensible, easy to deploy, easy to use, the advantages of task parallel, so it will be introduced to the collaborative filtering recommendation algorithm, the algorithm is distributed and parallel programming realization of acceleration, distributed collaborative filtering recommendation algorithm based on CPU. And analyzes the relative acceleration effect of single GPU and distributed CPU based on parallel processing.Based on the classic Movie Lens data sets, this paper discovered that the GPU and the distributed CPU have significant effect on accelerating the collaborative filtering recommendation algorithm, and further understand the working principle and practical significance.Due to the distance calculation is the core of many machine learning and data mining algorithms of computing tasks, the accumulation and summary of the CPU and GPU based parallel implementation experience accelerated distance calculation, which will have some reference to the improvement of other intelligence algorithms to improve the performance.