A method for weighting multi-valued features in content-based filtering

Abstract

Content-based recommender systems (CBRS) and collaborative filtering are the type of recommender systems most spread in the e-commerce arena. A CBRS works with two sets of information: (i) a set of features that describe the items to be recommended and (ii) a user's profile built from past choices that the user made over a subset of items. Based on these sets and on weighting items features the CBRS is able to recommend those items that better fits the user profile. Commonly, a CBRS deals with simple item features such as key words extracted from the item description applying a simple feature weighting model, based on the TF-IDF. However, this method does not obtain good results when features are assessed in multiple values and or domains. In this contribution we propose a higher level feature weighting method based on entropy and coefficients of correlation and contingency in order to improve the content-based filtering in settings with multi-valued features. © 2010 Springer-Verlag.