海角精品黑料

Yifan Feng

NUS Business School, Singapore

A Mallows-type Model for Preference Learning from Ranked Choices

Date: Friday, October 31, 2025
Time: 10:00 - 11:00 am
Location: Bronfman Building, Room 046

Abstract

The first part of the talk introduces a novel distance-based (鈥淢allows-type鈥�) ranking model built on a new distance function, the reverse major index (RMJ). The model is motivated by the problem of learning customer preferences from a general feedback structure we call "ranked choices," where each participant selects and ranks their top-k items from a displayed assortment. The RMJ-based ranking model yields surprisingly parsimonious, closed-form choice probability distributions over arbitrary display sets, which makes estimation both efficient and scalable in practice. We establish model identifiability and provide strong consistency guarantees for the estimator, even under model misspecification and low variation in assortments.

If time permits, the second part of the talk explores how (ranked) choice modeling can enhance LLM alignment. Current alignment methods typically rely on pairwise preference optimization. While simple, this approach overlooks richer forms of human feedback. To address this, we propose Ranked Choice Preference Optimization (RCPO). It is a unified framework that connects preference optimization with (ranked) choice modeling through maximum likelihood estimation. The framework is versatile, accommodating both utility-based and rank-based choice models. We embed the Mallows鈥揜MJ model into RCPO as an illustration, showing how appropriate feedback structures, combined with the right model, can yield more effective alignment.