SURE: Shift-aware, User-adaptive, Risk-controlled Recommendations

Nitin Bisht, Xiuwen Gong, Huan Huo, Guandong Xu

Keywords: Dynamic Prediction Sets

Abstract: Although Sequential Recommender Systems (SRS) have been well developed to capture temporal dynamics in user behavior, they face a critical gap in formal performance guarantees under preference shifts. When preferences change, predictions often become unreliable, undermining user trust and threatening long-term platform success. To address this challenge, we introduce **SURE** (**S**hift-aware, **U**ser-adaptive, **R**isk-controlled R**E**commendations), a dataset- and model-agnostic framework that provides adaptive recommendation sets with formal coverage guarantees while remaining compact under preference shifts. Specifically, SURE (i) ensures validity through a loss-based change-point mechanism that adaptively updates calibration thresholds upon detecting preference shift, (ii) maintains compact recommendation sets by stabilizing predictions with a Hedge-weighted ensemble of bootstrapped experts, preventing validity from degenerating into impractically large outputs, and (iii) guarantees robustness under non-stationarity by deriving finite-sample bounds that ensure the ensemble’s expected set size remains close to the best expert while controlling the utility-based risk in recommendation. Extensive experiments across multiple datasets and base models validate the effectiveness of the proposed framework, which aligns with our theoretical analysis.

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

Submission Number: 2903