EE ZOOM Seminar: Collaborative Preference Learning

https://tau-ac-il.zoom.us/j/86876252910?pwd=uhoOq5zjVquXvqFph3b3GcVlC5Pbq0.1
Meeting ID: 868 7625 2910
Passcode: 884113

Electrical Engineering Systems Seminar

Speaker: Tal Kravarusic

M.Sc. student under the supervision of Dr. Wasim Huleihel

Wednesday, 18th June 2025, at 16:00

Collaborative Preference Learning

Abstract

Preference learning from pairwise comparisons plays a central role in machine learning, with broad applications in recommendation systems, ranking tasks, and decision-making processes. In this paper, we study the problem of online preference learning in a challenging multi-user setting, where each user provides at most a single comparison for any pair of items. This setup significantly limits the available data, making standard averaging techniques inapplicable. To address this, we propose a collaborative learning framework that leverages the structural similarity among users. Each user is modeled by a latent pairwise preference matrix, from which Borda scores, quantifying the likability of each item, can be derived. These scores serve as a robust surrogate for full rankings, especially under noise or partial observations. We assume users belong to a small number of hidden types or rankings, which enables clustering and knowledge sharing across users. Under standard assumptions, such as non-ambiguity, incoherence between types, and strong stochastic transitivity, we introduce algorithms to recover Borda scores and identify top-ranked items. Our algorithms combine type-based clustering, empirical estimation, and noisy matrix completion to produce accurate inferences with provable guarantees. We provide two main recovery results: one under incoherence assumptions and one without, relying instead on ranking-level structure. Additionally, we propose a binary-search-style algorithm for efficiently identifying the top L items without recovering full rankings. In all cases, we derive bounds on the sample complexity required for successful inference. Our work contributes practical algorithms and theoretical insights for preference learning under highly constrained data, advancing the applicability of collaborative learning in real-world systems.

השתתפות בסמינר תיתן קרדיט שמיעה = עפ"י רישום בצ'ט של שם מלא + מספר ת.ז.