You Don’t Bring Me Flowers: Mitigating Unwanted Recommendations Through Conformal Risk Control
Автор: ACM RecSys
Загружено: 2025-10-02
Просмотров: 194
Giovanni examines how recommender systems can mitigate exposure to unwanted or harmful content. Using data from a large video platform that includes explicit “do not recommend” feedback, the team finds that short watch time predicts negative feedback and that users still re-encounter some reported videos. They model risk as the fraction of items a user would report and apply conformal risk control to choose a filtering threshold that guarantees the expected risk stays below a user-specified alpha. Because heavy filtering reduces item availability, they replace filtered items with previously liked, safe ones selected via initial watch time. Offline experiments show calibrated risk control, larger recommendation sets, and trade-offs in NDCG and serendipity. Careful selection of safe items is required to maintain guarantees.
Доступные форматы для скачивания:
Скачать видео mp4
-
Информация по загрузке:
