PrefPalette is a framework that considers not only the preferences themselves but also the reasons for them when understanding user preferences. Unlike existing preference models that treat human judgment as a black box, PrefPalette decomposes preferences into attribute dimensions and adjusts predictions to the values of the social community. Based on multi-attribute decision theory, it works by (1) generating synthetic training data to isolate the effects of individual attributes (e.g., formality, humor, cultural values), and (2) attention-based preference modeling to learn how the social community weights these attributes. When evaluated on 45 social communities on Reddit, it achieves an average prediction accuracy of 46.6% higher than GPT-4o. Academic communities show characteristic preference profiles for verbosity and provocation, conflict-oriented communities for sarcasm and directness, and support-oriented communities for empathy.
