BAHSD framework boosts recommendation accuracy by 80% for niche users

Sequential recommendation systems are often deployed as black-box APIs, making model extraction a popular approach to replicate their capabilities locally. However, the long-tail distribution of user interactions creates severe signal heterogeneity: dense head sequences cause the teacher model's preferences to solidify, biasing extraction toward local patterns, while sparse tail sequences yield flat, noisy predictions. Existing one-size-fits-all extraction methods overfit to noise and fail to transfer knowledge effectively.

To bridge this gap, the authors introduce BAHSD, a black-box adaptive distillation framework. It first employs a multi-scale consistency probing mechanism to implicitly quantify the reliability of each signal. Then, it applies an adaptive hierarchical objective: dynamic-temperature KL divergence mitigates preference solidification for high-confidence signals, while ranking consistency and InfoNCE contrastive learning provide noise-robust enhancement for low-confidence signals. In experiments, BAHSD consistently outperforms baselines, achieving up to 4.98% gain over the teacher model and over 80% improvement for tail users. This plug-and-play solution enables high-fidelity extraction from black-box recommendation APIs.