Improving Ranking-based Recommendation by Social Information and Negative Similarity

Recommender system is able to suggest items that are likely to be preferred by the user. Traditional recommendation algorithms use the predicted rating scores to represent the degree of user preference, called rating-based recommendation methods. Recently, ranking-based algorithms have been proposed and widely used, which use ranking to present the user preference rather than rating scores. In this paper, we propose two novel methods to overcome the weaknesses in VSRank, a state-of-the-art ranking-based algorithm. Firstly, a novel similarity measure is proposed to make better use of negative similarity; secondly, social network information is integrated into the model to smooth ranking. Experimental results on a publicly available dataset demonstrate that the proposed methods outperform the existing widely used ranking-based algorithms and rating-based algorithms considerably.