CosSimReg: An effective transfer learning method in social recommender system

Abstract

Traditional recommender systems perform poorly when training data is sparse. During past few years, researchers have proposed several social-based methods to alleviate this sparsity problem. The basic assumption of these social recommender systems is that friends should have similar interests. However, this assumption does not always hold due to the heterogeneity between recommendation domain and social domain. Thus, knowledge transferred from social network often contains noises. To solve this problem, in this paper, we analyze and identify what knowledge is useful during transfer learning process, and develop a method, called Cosine Similarity Regularization (CosSimReg), to transfer only useful information from social domain. CosSimReg is able to minimize the negative effects of noisy data in social network, thus improving the performance. Experiments on two real life datasets demonstrate that CosSimReg performs better than the state-of-the-art approaches. © 2014 Springer International Publishing Switzerland.