SimInter: Evidence-Grounded User Simulation with Interview-Based Profile Enrichment for Conversational Recommender Systems

Key points

• Conversational Recommender Systems (CRSs) require reliable user simulators for cost-effective evaluation.1
• LLM-based simulators have made realistic simulation possible but typically condition on profiles derived from explicit ratings, demographics, and static sources capturing surface preferences while omitting implicit behavioral characteristics such as decision-making style, aspect-level preferences, and domain expertise.1
• We instantiate and evaluate SimInter on movie recommendation.1
• We simulate 300 ReDial users and, on a matched subset of 100 users, compare SimInter with a statistics-only baseline: SimInter significantly increases the diversity of expressed preference aspects (filtered aspect entropy: 1.619 vs. 0.859; paired d = 0.96, Holm-adjusted p < 0.001).1

Sources (1)

• [1]SimInter: Evidence-Grounded User Simulation with Interview-Based Profile Enrichment for Conversational Recommender Systems
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 13, 12:52 PM

  We propose SimInter, a pipeline that elicits these attributes via synthetic interviews between two LLM agents, grounded in the user's statistical profile and historical conversational evidence.

  Conversational Recommender Systems (CRSs) require reliable user simulators for cost-effective evaluation.

Extractive summary: sentences quoted from the sources.