Extracting Implicit Friends from Heterogeneous Information Network for Social Recommendation

With the popularity of online social networks, social relation data is becoming increasingly important to alleviate the data sparsity and cold-start problem of the traditional recommender systems. Social relations, such as trust or friend relationships, are used as complement source to user feedback data (e.g. item rating). However, using explicitly issued social relations directly may generate sub-optional recommendation results because of the inherent drawbacks of explicit social relations. To address the inherent drawbacks of explicit social relation, we incorporate top-k implicit friends, who can be identified from a heterogeneous information network established by user feedback and user social relation data, into a matrix factorization method to make social recommendations. Experimental results on real-world datasets FilmTrust and Douban show that our method can improve the performance of rating prediction, compared to the social recommender systems using explicit social relation and non-social recommender system.