Multi-Label Bias-Based Predictor

A recommender system predicts the future preference of a set of items for a user by computing the similarity between users or items, and recommends the top items based on the prediction. Collaborative filtering, the most popular approach to build recommender systems, has been successfully employed in many commercial and non-commercial applications but has reached the limit of performance growth. A novel way to predict user preferences is needed to overcome this problem. Bias-based predictor (BBP) is a prediction model for recommender systems that assumes that bias and preference are highly correlated. In predictive models, bias generally means the intercept where our line intercepts the y-axis, offsetting all predictions we make. Users’ preference ratings are likely to be influenced by biases in many different aspects. This implies that bias can play more significant roles in preference prediction, rather than simply intercepting the y-axis. This paper proposes a prediction model for recommender systems that takes into account bias in multiple aspects. The proposed model called multi-label bias-based predictor (MLBBP) extends the conventional BBP to allow for a more in-depth analysis of bias. Through experiments with movie data, it was demonstrated that MLBBP performs better than BBP. The trained MLBPP model produces numerical data that can explain why a specific item is recommended to a user.