Accelerated Online Learning for Collaborative Filtering and Recommender Systems

Collaborative filtering (CF) is one of the major approaches to building recommender systems. Traditional batch-trained algorithms for CF suffer from some drawbacks, and online learning algorithms for CF, is a promising tool for attacking the large-scale dynamic problems. However, the low time complexity of online algorithm often be accompanied by low convergence rate, and the convergence rate of current dual-averaging online algorithm is only O(1/√T) up to T-th iteration. In order to tackle this problem, we propose a novel accelerated online learning framework for CF. Our algorithm has a accelerated capability, and its theoretical convergence rate bound is O(1/T²). Moreover, the proposed algorithm has low time and memory complexity, and scales linearly with the number of observed ratings. The experimental results on real-world datasets demonstrate the merits of the proposed online learning algorithm for large-scale dynamic collaborative filtering problems.