Online collaborative filtering via fusing the rating consistencies

Collaborative filtering (CF) is one of the key techniques used in online recommender systems, most traditional CF algorithms, however, are designed in the offline settings, hence are not suitable for the systems that need to be adjusted whenever new data arrive. Addressing to the problem, we propose Lefuse, an online collaborative filtering model. Our model is based on a set of identically and independently distributed assumptions on the user–item ratings, with the assumptions, we derive a counterpart set of consistency conditions, and formulate the prediction task as a constrained optimization problem. Specifically, under the Gaussian distribution setting, we derive the closed-form solution to the problem, which leads to a time and space efficient online CF algorithm. In the evaluations, we conduct comprehensive experiments with the proposed model and three other state-of-the-art online CF algorithms on four large scale real rating data sets, all results show the superiority of our algorithm.