Exploring GTRS Based Recommender Systems with Users of Different Rating Patterns

Abstract

Recommender systems predict a new user’s opinion on a collection of items by analyzing preference information of similar users. The Pawlak rough set (PRS) model is one of the effective tools to make personalized recommendations. The game-theoretic rough set (GTRS) model improves the quality of PRS based recommendations by determining a pair of thresholds that could achieve a tradeoff between two prominent recommendation evaluation metrics, accuracy and coverage. It should be noted that the performance of a recommendation algorithm may be affected by the rating patterns of the users in the considered dataset. The aim of this research is to evaluate how the performance of the PRS based and the GTRS based recommendations vary on user groups with different rating patterns. We conducted comparative experiments on five different data samples. The experimental results suggest that compared to the PRS model, the GTRS model could not only obtain an improvement in coverage level, but also achieve an equal accuracy level on each of the considered data samples. In particular, it achieved a bigger advantage over the PRS model on user groups that make a smaller number of rating records. This performance difference indicates that compared to the PRS model, the GTRS model is a better solution to make high quality personalized recommendations on small-scale datasets with fewer rating records stored in the database.

Keywords