Combining neighborhood-based and model-based on multi-criteria recommendation

Recommender system is a tool invented to filter information that seeks to provide personalized recommendations. The traditional recommender system makes the recommendations using the overall preferences toward items provided by the users. However, the multi-criteria recommender system suggests that the overall preferences of each individual user can be affected by his unequal personal interest in some criteria of the items. Learning such effect of each criterion becomes the key to produce more personalized recommendations. Most of methods in recommender systems are based on the neighborhood-based or the model-based techniques. To improve the performance of the recommendation, both techniques are often aggregated together. In this work, a novel multi-criteria recommendation technique is proposed. The prediction from each criterion is made by considering the trade-off between the neighborhood-based and the model-based techniques. The effects if the criterion ratings to the overall rating are measured by the similarities among the user preference patterns, extracted from matrix factorization. The overall rating is then predicted by weighted averaging the predictions from all criteria, using such criteria effects as the weights. The evaluation shows that our proposed method outperforms various well-known techniques on both single and multi-criteria recommendation.