Interactive Feedback Loop with Counterfactual Data Modification for Serendipity in a Recommendation System
Gyewon Jeon, Sangyeon Kim, Sangwon Lee
Abstract
Users often encounter tedious recommendations as they are continuously exposed to the recommendation system. In response to this issue, serendipity in a recommendation system has been introduced to generate novel and unexpected recommendations while keeping them relevant to users’ previous preferences. This study proposes an interactive feedback loop for a serendipity in a recommendation system that allows users to directly explore content via counterfactual manipulation of features. Specifically, users indicate their preferences through the “what-if” based customization of content meta-information, and these modifications influence their usage history, thereby enabling the elicitation of serendipitous items. To validate the proposed feedback loop, we conducted a scenario-based experiment and compared system-initiated and user-intervened recommendations. The results reveal that counterfactual exploration can help to generate serendipitous recommendations. This study contributes to providing a user-friendly recommendation system that can retrieve preference-reflected recommendations through user interaction.