Kernel meets recommender systems: A multi-kernel interpolation for matrix completion

摘要

A primary research direction for recommender systems is matrix completion, which attempts to recover the missing values in a user-item rating matrix. There are numerous approaches for rating tasks, which are mainly classified into latent factor models and neighborhood-based models. Most neighborhood-based models seek similar neighbors by computing similarities in the original data space for final predictions. In this paper, we propose a new neighborhood-based interpolation model with a kernelized matrix completion framework, with the impact weights provided by neighbors computed in a new Hilbert space containing more features. In our model, the kernel function is combined with a similarity measurement to achieve better approximation for unknown ratings. Furthermore, we extend our model with a non-linear multi-kernel framework which learns weights automatically to improve the model. Finally, we conduct extensive experiments on several real-world datasets. The outcomes show that the proposed methods work effectively and improve the performance of the rating prediction task compared to both the traditional and state-of-the-art approaches.