Unsupervised Expert Finding in Social Network for Personalized Recommendation

Personalized Recommendation has drawn greater attention in academia and industry as it can help people filter out massive useless information. Several existing recommender techniques exploit social connections, i.e., friends or trust relations as auxiliary information to improve recommendation accuracy. However, opinion leaders in each circle tend to have greater impact on recommendation than those of friends with different tastes. So we devise two unsupervised methods to identify opinion leaders that are defined as experts. In this paper, we incorporate the influence of experts into circle-based personalized recommendation. Specifically, we first build explicit and implicit social networks by utilizing users' friendships and similarity respectively. Then we identify experts on both social networks. Further, we propose a circle-based personalized recommendation approach via fusing experts' influences into matrix factorization technique. Extensive experiments conducted on two datasets demonstrate that our approach outperforms existing methods, particularly on handing cold-start problem.