Users’ preference-degree considered diffusion for recommendation on bipartite networks

In recommender systems, ratings represent the degree of users’ preference. However, in the top-L recommendation, negative ratings are simply discarded or different ratings are regarded as the same. In order to evaluate the ability to recommend objects with high ratings, we put forward a metric called Average High-score Ratio (AHR) to compute the average ratio of ratings of recommended objects to the maximum score allowed by the system. The higher of AHR, the more objects with high ratings will be recommended to users. Computing AHR does not need to collect extra information except ratings, and it is a proper metric to measure user satisfaction. We also propose users’ preference-degree considered diffusion algorithm for recommendation, which distinguishes different ratings and is parameter-free. Compared with some classic and recent proposed methods, users’ preference-degree considered diffusion algorithm has the best performance on recommendation accuracy and user satisfaction, and it gets the second-best performance on Hamming distance, novelty, and coverage, only next to the Heat Conduction method. Our work gives exploration on designing recommendation algorithms with better user satisfaction.