GPU-based Parallel Collaborative Filtering Algorithm With Application

Item-based collaborative filtering is one of the most famous algorithms in therecommender system field. But its shortcoming will turn out when there are millionsof users and millions of items, which is very normal in the product environment. Andthis becomes the main obstacle of its application. People use clustering, dimensionreduction, and computer cluster technology to solve this problem of big data. Butdimension reduction will reduce the quality of the dataset and computer clustertechnology not only requires high web quantity and stability, but also needs expensivehardware, and the code is hard to write too, so that the computer cluster is not broadlyused in medium-sized and small enterprises.In order to improve the scalability of collaborative filtering algorithm, this paperprovides an algorithm of GPU based parallel collaborative filtering. GPU uses singleinstruction multiple data (SIMD), and is proper for calculating weak logic and largeamount data, which is exactly the features of collaborative filtering. This paperdesigns the whole procedure of the GPU based parallel collaborative filtering andimplements it using CUDA. Our experimental results show our algorithm is veryefficient to process the large-scale datasets with good accuracy. Then based on thisalgorithm, we analyze the application in network recruitment for recommendingpositions to the website users. Finally, we present a parallel position recommendersystem for network recruitment.There are three contributions. Firstly, we transform the core computation ofcollaborative filtering into matrix computation. Secondly, we use GPU to archive thegoal of high performance, low cost, and low power consumption. Thirdly, we applythis algorithm to solve the recommendation problem in network recruitment.