Effects of negative ratings on personalized recommendation
Wei Zeng, Mingsheng Shang · 2010
In most recommender systems, the active user's preferences can be denoted by multi-graded rating data (1 to 5 in MovieLens etc.). When using the available ratings, some recommendation algorithms transfer multi-graded rating data into binary rating data ignoring the actual value of ratings, while some others just use positive ratings (no smaller than 3 in 1-5 rating structure e.g.) to recommend items for users. In the former case, positive ratings and negative ratings are treated equally while in the latter case the negative ratings are not used at all. In this paper, we use a tunable parameter to combine the positive ratings and the negative ratings, and a diffusion-based process in a weighted bipartite networks to make personal recommendation. We test the proposed method with three data-sets. The results demonstrate that the negative ratings can help to improve the accuracy of recommendation algorithm.