A multicriteria recommendation method for data with missing rating scores

This paper proposes a recommendation method for multi-criteria (MC) collaborative filtering, where users are required to give rating scores from multiple aspects to each item and systems utilize the rich information to improve the recommendation accuracy. One drawback of MC recommender systems is user's cost to give scores to items because it requires rating scores on MC for each item. To overcome this drawback, we aim at developing a MC recommender system that allows missing rating information. This paper proposes generative models for MC recommendation that are robust against missing scores. In these models we convert a list of rating scores on MC to a low dimensional feature space. Correlation among scores on MC is embedded in the feature space. So we can expect that a score list is mapped to a close point in the feature space even if some scores are missing. We conducted experiments to check the robustness of the proposed models by using Yahoo! movie data and experimentally show that they are less affected by missing information compared to Pearson correlation base collaborative filtering method.