Problem
Flat interaction sequences hide session structure. Dense attention can also spend computation on irrelevant history, weakening useful preference signals.
Approach
HPGR first learns hierarchical item representations through session-based masked-item pretraining. Preference-aware fine-tuning then uses sparse attention to focus on relevant historical items.
Public evidence
The paper reports evaluation on proprietary APPGallery data and an online A/B test, comparing with HSTU and MTGR. These settings support the reported efficiency and recommendation findings; they are not a public-benchmark replication.
Takeaway
Session structure and selective attention provide complementary controls over what a generative recommender learns from its history. Generalization beyond the reported industrial setting requires evaluation.
Sources
Bibliographic details
- Authors
- Zerui Chen, Heng Chang, Tianying Liu, Chuantian Zhou, Yi Cao, Jiandong Ding, Ming Liu, Bing Qin
- Venue
- WWW 2026
- Status
- Published