PrefPalette: Personalized Preference Modeling with Latent Attributes
Shuyue Stella Li, Melanie Sclar, Hunter Lang, Ansong Ni, Jacqueline He, Puxin Xu, Andrew Cohen, Chan Young Park, Yulia Tsvetkov, Asli Celikyilmaz
2025-07-23
Summary
This paper talks about PrefPalette, a new AI system that breaks down people's preferences into different qualities or traits and uses attention-based methods to understand how different social groups value these traits.
What's the problem?
Most existing AI models treat what people like as a mystery without explaining why they prefer certain things, making personalization limited and hard to interpret.
What's the solution?
The researchers created PrefPalette, which generates synthetic training examples to isolate the effects of individual traits like humor or politeness, and then uses a model that pays attention to how different communities weigh these traits to predict preferences more accurately and clearly.
Why it matters?
This matters because PrefPalette not only predicts what people like better than top models but also explains why, helping build AI systems that respect and reflect the values of different social groups, making personalization more trustworthy and meaningful.
Abstract
PrefPalette decomposes preferences into attribute dimensions and models them with attention-based techniques, outperforming GPT-4o in prediction accuracy and providing interpretable insights into community-specific values.