An Improvement of Recommender Systems Considering Preference Inconsistencies Based on Interrelationship Mining

Abstract

Abstract In this paper, we aim to improve recommendation systems by utilizing inconsistencies in human preferences, which often occur, as meaningful features through interrelationship mining. A previous study by Kuroda proposed a system that presents users with pairs of tourist images and uses any resulting contradictions in their intuitive choices as preference characteristics. However, the experimental results showed that no such inconsistencies actually occurred. To address this issue, we propose a method to design questions that are more likely to elicit preference inconsistencies. We conduct experiments to evaluate how frequently such inconsistencies appear in user choices. Furthermore, we compare the performance of our method with conventional approaches to examine how the presence or absence of inconsistencies affects recommendation accuracy.