RAGR boosts recommendations by blending user reviews directly into AI sequences
New method from 11 researchers uses reviews as tokens for better generative recommendations.
A team of 11 researchers from multiple institutions has introduced RAGR (Review-Augmented Generative Recommendation), a new framework that fundamentally changes how user reviews are used in sequential recommendation systems. Traditional generative recommendation (GR) methods model user behavior purely through item interaction sequences, missing the latent evaluative factors that reviews reveal — such as why a user chose one product over another. RAGR addresses this by constructing a mixed behavioral-semantic sequence: it interleaves item semantic IDs and review semantic IDs in chronological order, allowing review signals to participate directly in autoregressive next-token generation.
To ensure the model remains focused on the core recommendation task, RAGR employs an Item-Centric Task Generation Alignment strategy based on direct preference optimization (DPO). This encourages the model to favor predicting item tokens over review tokens at prediction positions, preventing the reviews from diluting the recommendation objective. Experiments on three real-world datasets (likely including Amazon or Yelp-style data) show consistent and significant improvements over strong GR backbones across all evaluation metrics. The code and data are publicly available, marking a practical step toward more context-aware generative recommenders.
- RAGR interleaves item and review semantic IDs chronologically in a single generative sequence, unlike prior GR methods.
- Uses direct preference optimization (DPO) to bias the model toward item token prediction for recommendation accuracy.
- Achieves consistent gains across three real-world datasets vs. strong generative baselines.
Why It Matters
Reviews are abundant but underused in generative recommenders; RAGR turns them into actionable signals for better personalized suggestions.