Exploiting Nonlinear Relationships for Top-N Recommender Systems

To alleviate the information overload problem, recommendation technology has emerged and flourished. %As a most widely used recommendation technique, collaborative filtering algorithms suffer from data sparseness and cold start problems. Consequently, it is difficult to obtain accurate similarities between users and items and reliable basis of the predictions with these algorithms, leading to sub-optimal recommendation quality. Many state-of-the-art methods usually assume that the data is distributed on a linear hyperplane, which is not the case. The rating data reflect the many-sided interests of users and usually have nonlinear dependencies. In this paper, we map the data into a higher dimensional space and learn the similarity information in this new feature space. Kernel methods are known to be effective for capturing the complex relations in many real world applications. In the first place, a single kernel based algorithm is proposed. It is known that the performance of kernel methods is largely dependent on the choice of kernel. To alleviate such a dependence, we further develop a multiple kernel based algorithm. Experimental results on six real world datasets demonstrate that the proposed algorithms significantly improve the performance of several state-of-the-art recommendation methods.