RAGR: Review-Augmented Generative Recommendation

2026-05-17

AI Summary

This work addresses a structural limitation in existing generative recommendation approaches, which model only user interaction sequences while neglecting the decision-making rationale embedded in reviews. To overcome this, we propose a novel generative recommendation framework that treats reviews as first-class elements by integrating them into user behavior sequences, forming hybrid sequences of interleaved items and reviews. Within a unified token space, the model jointly learns semantic IDs and performs autoregressive decoding. To steer the generation toward accurate item prediction, we introduce an item-centric task alignment strategy based on Direct Preference Optimization (DPO). Extensive experiments on three real-world datasets demonstrate that our method consistently and significantly outperforms strong baselines across multiple evaluation metrics.

Abstract

Sequential recommendation (SR) is traditionally formulated as next-item prediction over a chronological sequence of interacted items. Although recent generative recommendation (GR) methods introduce new machinery, such as semantic IDs, autoregressive decoding, and unified token spaces, they largely inherit the same item-only modeling assumption. We argue that this design constitutes a structural bottleneck, because user decision-making is not purely behavioral: while item interactions reveal what users choose, review feedback often explain why they choose it by exposing latent evaluative factors. Motivated by this observation, we propose Review-Augmented Generative Recommendation (RAGR), a novel GR framework that incorporates review feedback directly into the generative user sequence rather than treating reviews as auxiliary side information. Specifically, RAGR introduces a Review-Augmented User Sequence Modeling mechanism that interleaves item semantic IDs and review semantic IDs in chronological order to construct a mixed behavioral-semantic sequence, enabling review signals to participate directly in autoregressive next-token generation. To preserve the recommendation objective, we further introduce an Item-Centric Task Generation Alignment strategy based on direct preference optimization (DPO), which encourages the model to favor item tokens over review tokens at prediction positions. Experiments on three real-world datasets show that RAGR yields consistent and significant gains over strong GR backbones across all metrics. Our code and data are available at \url{https://github.com/Zhang-Yingyi/TKDE_RAGR}.

Problem

• sequential recommendation
• generative recommendation
• review feedback
• user decision-making
• item-only modeling

Innovation

• generative recommendation
• review augmentation
• semantic ID
• autoregressive decoding
• direct preference optimization