Research On Collaborative Filtering Algorithms Based On Graph Spectral Transfer

As one of the most effective and popular approaches in recommender systems, collaborative filtering(CF) makes personalized recommendations for users by dealing with a user-item rating matrix. Since CF focused on managing numerical information, it has many advantages compared to other recommendation systems. Whereas, there are also many bottlenecks in CF, typically including data sparsity, code start problem, scalability etc.. Our work aims at the data sparsity. Transfer learning aims to alleviate the information deficiency in collaborative filtering, by transferring auxiliary information into the target domain. Most previous works about transfer learning have rigid assumptions about the auxiliary data which have been barriers for their practical application. For example, the auxiliary domains should share the identical users or items with the target domain, and the dimensions of latent features extracted from the auxiliary domain should be equal to the target domain’s. Mostly, some of these methods use multiple tradeoff parameters that should be carefully chosen via searching. In this paper, our work focuses on the above restrictions, which can be devided into two parts:(1) Aimed at the consistence problem, we propose a novel transfer learning method, named Feature-Based Graph Transfer for Collaborative Filtering(FBGT). In this algorithm, based on Nonnegative Matrix Factorization(NMF), we predict the user/item’s preference by transferring their neighbor information from auxiliary domains. Since we only required the users or items in the auxiliary domains are partly overlap with the target domain, so there is no rigid limitation on the size of auxiliary data. In our model, we utilize the intrinsic structures of users and items from two auxiliary domains respectively, instead of utilizing their ratings directly, so we do not require the feature dimension in accordance with the target domain’s. We also design two ways to construct the graphical structures in the auxiliary domains. The numerical experiments on real datasets prove the improvement of FBGT.(2) Based on FBGT, we propose another transfer method in the consideration of that FBGT is not a convex model. The graph regularized matrix factorization algorithm can inherit the advantages of FBGT, in which we combine the Regularized Matrix Factorization(RMF) model with graphical structure. And also a adaptive strategy is adopted in GRMF to reduce the parameter requirement. This algorithm utilizes the sparse and low-rank structures sufficiently. We design several numeral experiments on real-world data to show the improvement of our proposed algorithms.