GPU accelerated item-based collaborative filtering for big-data applications

Recommendation systems are a popular marketing strategy for online service providers. These systems predict a customer´s future preferences from the past behaviors of that customer and the other customers. Most of the popular online stores process millions of transactions per day; therefore, providing quick and quality recommendations using the large amount of data collected from past transactions can be challenging. Parallel processing power of GPUs can be used to accelerate the recommendation process. However, the amount of memory available on a GPU card is limited; thus, a number of passes may be required to completely process a large-scale dataset. This paper proposes two parallel, item-based recommendation algorithms implemented using the CUDA platform. Considering the high sparsity of the user-item data, we utilize two compression techniques to reduce the required number of passes and increase the speedup. The experimental results on synthetic and real-world datasets show that our algorithms outperform the respective CPU implementations and also the naïve GPU implementation which does not use compression.