Recommender systems are becoming very useful for competitive businesses. It is very important for recommender systems to extract user preferences accurately by utilizing logs that record user behavior. Furthermore, user behavior should be analyzed from multiple aspects, storing the results as multicriteria rating scores. If the rating information is sparse, then systems are forced to compensate. One way to treat sparseness is to use a latent model that maps users and items to a small number of groups. To predict rating scores from such a model, we need to aggregate the data appropriately. This paper proposes a method for combining a latent model with a proposed regression technique. We evaluated the proposed method for the Yahoo! Movie data set and show empirically that the proposed combination improves the recommendation accuracy.