Yujin Ju

Sequential recommendation aims to predict the next item from a user’s past interaction sequence, and has largely advanced around how accurately the ground-truth item is ranked. In practice, however, a recommendation list influences subsequent user behavior and item exposure, forming a closed-loop feedback process. As this process repeats it can lead to hot-item concentration and coverage collapse, which are difficult to capture by ranking accuracy alone. To address this problem we propose CLEREC, a closed-loop exposure-aware sequential recommendation framework. CLEREC separates a preference path from an exposure-aware path and integrates the exposure-aware information as a gated residual correction through a preference-conditioned fusion gate. On the H&M and Amazon Electronics datasets, CLEREC improves Hit@10 by 32.1% and 15.8%, respectively, over the no-prior setting. Moreover, by adjusting the strength of the popularity prior, it raises ranking accuracy while maintaining coverage of 0.359 on H&M and 0.427 on Electronics, thereby mitigating concentration of hot item. These results show that accuracy and exposure-related behavior should be evaluated jointly in closed-loop sequential recommendation and that CLEREC can serve to adjust the operating point between accuracy and coverage in recommendation settings where repeated item exposure matters.