Dissimilarity Features in Recommender Systems

In the context of recommenders, providing suitable suggestions requires an effective content analysis where information for items, in the form of features, can play a significant role. Many recommenders suffer from the absence of indicative features capable of capturing precisely the users´ preferences which constitutes a vital requirement for a successful recommendation technique. Aiming to overcome such limitations, we introduce a framework through which we extract dissimilarity features based on differences in preferences of items´ attributes among users. We enrich the representations of items with the extracted features for the purpose of increasing the ability of a recommender to highlight the preferred items. In this direction, we incorporate the dissimilarity features into different types of classifiers/recommenders (C4.5 and lib-SVM) and evaluate their importance in terms of precision and relevance. Experimentation on real data (Yahoo! Music Social Network) indicates that the inclusion of the proposed features improves the classifiers´ performance, and subsequently the provided recommendations.