Beyond the Flat Sequence: Hierarchical and Preference-Aware Generative Recommendations

A research team led by Zerui Chen and Heng Chang has introduced HPGR (Hierarchical and Preference-aware Generative Recommender), a breakthrough framework that addresses fundamental limitations in current generative recommendation systems. Published in ACM Web Conference 2026 and detailed in arXiv:2603.00980, HPGR challenges the "flat-sequence" assumption that treats user interactions as simple chronological lists, arguing this approach misses crucial hierarchical structures in real-world behavior patterns. The researchers identified two key problems with existing models like HSTU (Hierarchical Sequential Transduction Unit): they fail to capture session-based engagement hierarchies, and their dense attention mechanisms introduce significant noise when processing semantically sparse user histories.

The HPGR framework implements a sophisticated two-stage architecture that injects structural priors directly into the modeling process. First, a structure-aware pre-training stage employs session-based Masked Item Modeling (MIM) to learn hierarchically-informed item representations that capture semantic relationships. Second, a preference-aware fine-tuning stage leverages these representations to implement Preference-Guided Sparse Attention, which dynamically focuses computation only on the most relevant historical items. This dual approach reduces computational noise while improving signal-to-noise ratio. Empirical validation on Huawei's large-scale APPGallery dataset and online A/B testing demonstrated state-of-the-art performance, with HPGR outperforming strong baselines including HSTU and MTGR across multiple metrics. The framework represents a significant advance toward more efficient, accurate, and interpretable AI-powered recommendation systems.