A STUDY ON IMPLICIT FEEDBACK IN MULTICRITERIA E-COMMERCE RECOMMENDER SYSTEM

Recommender systems are personalized intelligent systems capable of helping people to easily locate their relevant information through recommendations from a large repository of information. In order to provide personalized recommendations, the accurate modeling of user's preferences is required. Modeling of user's preferences needs their relevance feedback on the recommendations. The relevance feedback may be collected either explicitly or implicitly. The explicit relevance feedback introduces intrusiveness problem whereas the implicit feedback can be inferred from normal user-system interactions without disturbing the user. The users expect accuracy in recommendations and effortless assistance from the recommender systems. The multicriteria user preference ratings are useful to improve the accuracy of recommendations. However, collecting multiple ratings increases the cognitive load of the user. We believe that a combined, implicit relevance feedback and multicriteria user preference ratings, approach improve the accuracy in recommendations and eliminate the intrusiveness problem of recommender systems. In order to fulfill the above needs and to better understand the potential behind the implicit relevance feedback approach under multicriteria ratings context, this study focuses a new implicit-multicriteria combined recommendation approach. A Music recommender system is developed for this experiment to evaluate the recommendation accuracy of implicit and explicit feedback approaches under the user-based and item-based prediction algorithms against different data sparsity levels, training/test ratio and neighborhood sizes. Out of this experiment, the implicit ratings based prediction algorithms provide better performance than the explicit ratings based prediction algorithms with respect to all the three sensitive parameters. It is also observed that the proposed IB_PIR prediction algorithm computes better predictions than other prediction algorithms. Finally, we discuss the study's implications for theory and practice and conclude with many suggestions for future research on non-intrusive, multicriteria recommender systems.