Guaranteed Top-Adaptive-K in Recommendation

Nitin Bisht, Xiuwen Gong, Guandong Xu

Keywords: User Personalized Recommendations

Abstract: Recommender systems (RS) are crucial in offering personalized suggestions tailored to user preferences. While conventionally, Top-\(K\) recommendation approach is widely adopted, its reliance on fixed recommendation sizes overlooks the diverse needs of users, leading to some relevant items not being recommended or vice versa. While recent work has made progress, they determine \(K\) by searching over all possible recommendation sizes for each user during inference. In real-world scenarios, with large datasets and numerous users with diverse and extensive preferences, this process becomes computationally impractical. Moreover, there is no theoretical guarantee of improved performance with the personalized K. In this paper, we propose a novel framework, **K-Adapt**, which determines dynamic K-prediction set size for each user efficiently and effectively. Specifically, it reformulates adaptive Top-\(K\) recommendation as a utility-based risk control problem, where a calibrated threshold based on user utility metrics determines the prediction sets. A lightweight greedy optimization algorithm efficiently learns this threshold to generate dynamic recommendations. Theoretical analysis is provided by establishing upper bounds on expected risk as well as near-optimality and stability of the learned threshold. Extensive experiments on multiple datasets demonstrate that the K-Adapt framework outperforms baseline methods in both performance and time efficiency, offering a guaranteed solution to fixed Top-\(K\) challenges.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 15777