ReFORM: Review-aggregated Profile Generation via LLM with Multi-Factor Attention for Restaurant Recommendation
New framework analyzes user reviews to highlight specific factors like ambiance or service, beating state-of-the-art models.
A team of researchers has introduced ReFORM, a novel AI framework designed to revolutionize how we get restaurant recommendations. The system addresses a key flaw in current LLM-enhanced recommenders, which often rely too heavily on basic item titles and the model's internal knowledge, ignoring the rich, decision-influencing details found in user reviews. ReFORM first uses a Large Language Model to meticulously parse restaurant reviews, generating factor-specific profiles for both users (what they care about) and items (how they are evaluated).
The core innovation is its 'Multi-Factor Attention' mechanism. This component analyzes the generated profiles to dynamically identify and weight the specific factors—such as service speed, price value, noise level, or dietary options—that are most critical to an individual user's decision-making process for a given context. This moves beyond one-size-fits-all scoring to a nuanced, personalized ranking.
In experiments on two restaurant datasets of different scales, ReFORM demonstrated superior performance over state-of-the-art baseline models. The researchers also conducted in-depth analyses to validate that the proposed modules effectively capture the sources of personalization, making the recommendation process more transparent and robust. The team has made their source code and datasets publicly available, encouraging further development in the field.
- Uses LLMs to create detailed user/item profiles from review text, not just titles.
- Employs a 'Multi-Factor Attention' mechanism to personalize recommendations based on what matters most to each user.
- Outperformed existing state-of-the-art models in tests on two real-world restaurant datasets.
Why It Matters
Moves recommendations beyond simple ratings to nuanced, review-driven personalization, promising better results for apps like Yelp or OpenTable.