On Top-k Recommendation using Social Networks

Recommendation accuracy can be significantly improved by incorporating trust relationships derived from social networks into collaborative filtering approaches. While there has been recent work on minimizing the root mean square error (RMSE), the top-k hit ratio is a more realistic accuracy measure in commercial systems, as only a small number k of items can be recommended to a user at a time. In this paper, we modify the training of various Matrix Factorization (MF) and Nearest Neighbor (NN) models as to maximize the hit ratio. This results in considerable improvements over the existing approaches in our experiments on two publicly available data sets. Surprisingly, we also found that the technical approach for combining feedback data (e.g. ratings) with social network information that works best for minimizing RMSE works poorly for maximizing the hit ratio, and vice versa.