A GPU-oriented online recommendation algorithm for efficient processing of time-varying continuous data streams

摘要

Research on recommendation systems has gained a considerable amount of attention over the past decade as the number of online users and online contents continue to grow at an exponential rate. With the evolution of the social web, people generate and consume data in real time using online services such as Twitter, Facebook, and web news portals. With the rapidly growing online community, web-based retail systems and social media sites have to process several millions of user requests per day. Generating quality recommendations using this vast amount of data is itself a very challenging task. Nevertheless, opposed to the web-based retailers such as Amazon and Netflix, the above-mentioned social networking sites have to face an additional challenge when generating recommendations as their contents are very rapidly changing. Therefore, providing fresh information in the least amount of time is a major objective of such recommender systems. Although collaborative filtering is a widely used technique in recommendation systems, generating the recommendation model using this approach is a costly task, and often done offline. Hence, it is difficult to use collaborative filtering in the presence of dynamically changing contents, as such systems require frequent updates to the recommendation model to maintain the accuracy and the freshness of the recommendations. Parallel processing power of graphic processing units (GPUs) can be used to process large volumes of data with dynamically changing contents in real time, and accelerate the recommendation process for social media data streams. In this paper, we address the issue of rapidly changing contents, and propose a parallel on-the-fly collaborative filtering algorithm using GPUs to facilitate frequent updates to the recommendations model. We use a hybrid similarity calculation method by combining the item-item collaborative filtering with item category information and temporal information. The experimental results on real-world datasets show that the proposed algorithm outperformed several existing online CF algorithms in terms of accuracy, memory consumption, and runtime. It was also observed that the proposed algorithm scaled well with the data rate and the data volume, and generated recommendations in a timely manner.