LLM-Assisted Debiasing for Rec Eval

_reachsumit · x · 2026-07-14

Using LLMs to Elicit User Preferences for Better Offline Rec Eval

This work proposes a method that uses an LLM to elicit user preference profiles, which are then used to infer missing relevance labels in offline Top-N recommendation evaluation, thereby mitigating popularity bias.

The primary goal isn't to improve the recommendation model itself, but to make offline evaluation better reflect real user preferences, reducing systematic errors caused by incomplete labels.

Original post →

More from Research

Research channel →