A new recommender algorithm on signed networks

Many real-world systems display opposite relationships and can be depicted as signed networks to study. On signed networks, positive/negative edges mean users like/dislike objects. This information is valuable and should be considered into recommendations. In this paper, we mainly study recommendations on signed networks that contain users’ purchase behaviors as well as attitude information, which not only can validate the accuracy of recommendation algorithms but also measure the users’ satisfaction degree after purchasing. Accordingly, we proposed a new recommender algorithm by defining an index P. We further compared our method to other four classical algorithms on three disparate datasets. The results show the accuracy of our method improves at most three times higher than other classic algorithms on recommending negative edges. In addition, the recommendation diversity of our method performs better than heat conduction algorithm which is generally recognized as an effective algorithm in terms of diversity. For instance, the value of Novelty dropped from 19.74 to 3.04 when comparing the heat conduction algorithm with our method on the Movielens dataset. In a word, our method can recommend the objects that are novel to users and ensure users’ satisfaction after purchasing.