Investigating Preference Elicitation Strategies in Conversational Recommender Systems

Abstract

This research develops theoretical and methodological foundations for modular, mixed-initiative conversational recommender systems (CRS). It investigates how such systems can be conceptualized and systematically evaluated. The aim is to improve preference elicitation, dialogue management and intent detection by investigating the role of large language models (LLMs) in these tasks. First, a literature review is provided to conceptualize the use of generative models in CRS, and behavioral archetypes in user sessions that inform system design are identified. A user study compares a traditional faceted search interface with system-driven CRS to assess trust, usability and preference extraction quality in order to assess system acceptance. The outlined work extends this prototype to support mixed-initiative dialogues, with components for intent detection, question generation and state-tracking and focuses on a modular evaluation, often missing in the current literature. The project aims to deliver a holistic CRS designed and evaluated with consultancy of end users and platform stakeholders.