Empirical Evaluation of Alternatives to the Aggregation of Neighbor’ Ratings in Social Filtering

Classical collaborative filtering algorithms that generate recommendations on the basis of item ratings use standard aggregation schemes that are devised a priori, without a process of empirical assessment. Nonetheless, common empirical evaluation procedures can be used to obtain evidence of the appropriateness of changing the behavior of the aggregation schemes. This paper reports the empirical evaluation of different alternatives in the aggregation of the rating of neighbors as considered in the classical Pearson formula for social filtering used in the original GroupLens system. The results point out to significant differences that may be hypothetized to be better models for the modeling of the “word-of-mouth” approach underlying collaborative filtering techniques.