추천 시스템의 스테디셀러 아이템: 리랭킹에 따른 노출 재분배와 추천 정확도 분석

초록

Recommender systems help users identify relevant items, but recommendation exposure can concentrate on a small set of popular items. Persistently popular items may also remain highly exposed in recommendation lists, motivating an examination of whether reducing their exposure creates recommendation opportunities for less popular items. We define these persistently popular items as steady seller items and analyze exposure redistribution and recommendation accuracy under re-ranking using music and movie datasets. We compare penalties restricted to steady sellers or equally sized item sets selected by cumulative interaction counts with a catalog-wide popularity penalty. We also vary the maximum number of steady sellers allowed in each recommendation list and the item groups prioritized for replacement. The results show that penalties restricted to steady sellers primarily redistribute exposure to other popular items, whereas exposure gains for less popular items depend on the scope of the penalty. Accuracy losses also vary with the number of steady sellers retained and the choice of replacement items. These findings indicate that reducing exposure for a particular popular group alone is insufficient to assess popularity bias mitigation. This study provides an empirical basis for jointly evaluating exposure reductions, the groups receiving redistributed exposure, and the associated accuracy costs.

키워드