CosSimReg: An Effective Transfer Learning Method in Social Recommender System

摘要: 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 that friends should similar interests. However, does not always hold due the heterogeneity between recommendation domain and domain. Thus, knowledge transferred from network often contains noises. To solve problem, in paper, we analyze identify what useful during transfer learning process, develop a method, called Cosine Similarity Regularization (CosSimReg), only information CosSimReg able minimize negative effects noisy network, thus improving performance. Experiments on two real life datasets demonstrate performs better than state-of-the-art approaches.