An Explanation Engine for Increasing Self-awareness on Movie Preferences
金 琢奇（2021年度修了）
Recently researches on explainable recommendation systems have attracted much attention The explainable recommendation system is to give explanatiOns about t he recommended items The explanations are designed to help users have a better understanding of the relationship between the items and the user 's preferences. Inspired by these studies our research fbcused on generating explanations to increase llsers' selfLawareness about their prefbrences.
We conducted our researchin the domain of movies. We designed a chat-based conversational system to commUnicate with the uSers and extract their prefbrences on movies・We generated feature-level personalized explanations based on the movie prefbrences・The explanations were divided into two Parts・First , we explained to the llSers abollt the feature values they prefer. Our approach fbr generating feature-level preferences was based on the frequency of movies containing these features in the user 's faNﾉorite movies. Second、we explained tO the users that whether their preferences were special or common The personalized explanations were then shown to the users・In the baseline system, we only generated common explanations The common explanations also had two parts. First , we randomly generated feature values and told the users that they preferred them・Second, we explained to the users that whether those feature valueS were special or common.
To evaluate the effectiveness of the explanations,we perfOrmed a user study to analyze the changes in user behaviors The users were supposed to give out five of their favorite movies and then score them. The system would give out explanations about their preferences. i.e.,The experiment system gave out personalized explanations and the baseline system gave Out common explanations The userS were supposed to score the five movies again after reading the explanations. We determined whether they had increased their selfLawareness based on how much they changed their scores，We fbund that our proposed system increased users selfawareness on movie preferences compared to the baseline system.
In future work, we will apply machine learning methods to calculate fbature-level preferences based on movie preferences.