Empirical Assessment of a Collaborative Filtering Algorithm Based on OWA Operators

Classical collaborative filtering algorithms generate recommendations on the basis of ratings provided by users that express their subjective preference on concrete items. The correlation of ratings is used in such schemes as an implicit measure of common interest between users, that is used to predict ratings, so that these ratings determine recommendations. The common formulae used for the computation of predicted ratings use standard weighted averaging schemes as the fixed aggregation mechanism that determines the result of the prediction. Nonetheless, the surrounding context of these rating systems suggest that an approach considering a degree of group consensus in the aggregation process may better capture the essence of the word-of-mouth philosophy of such systems. This paper reports on the empirical evaluation of such an alternative approach in which OWA operators with different properties are tested against a dataset to search for the better empirical adjustment. The resulting algorithm can be considered as a generalization of the original Pearson formula based algorithm that allows for the fitting of the aggregation behavior to concrete databases of ratings. The results show that for the particular context studied, higher orness degrees reduce overall error measures, especially for high ratings, which are more relevant in recommendation settings. The adjustment procedure can be used as a general-purpose method for the empirical fit of the behavior of collaborative filtering systems. (C) 2008 Wiley Periodicals, Inc.