Turning Cold Entities Warm: LLM-Based User and Item Interaction Simulation for Recommendations

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Recommendation throughout the cold-start period remains a challenge in recommender systems. Strictly cold entities (users/items) have no interactions, while warm-up entities only have a few unreliable interactions. Strict cold-start methods typically overlook warm-up entities, while warm-up methods are generally not designed for entities with no observed interactions. In this paper, we propose ProbLLM , an LLM-driven framework that reformulates cold-start recommendation as interaction simulation through a shared retrieve–score–select interface. LLM-based content retrieval first constructs a bounded candidate pool, after which a probability-form prompt elicits a bounded confidence score for every retained pair. Selected interactions retain these confidence scores as soft weights for downstream representation learning. Within this architecture, Filtering Simulation is followed by Refining Simulation for every cold target. Pairwise refinement elicits an LLM confidence score for each filtered candidate before pseudo-interaction selection. A two-phase sequential procedure first initializes the collaborative embeddings on observed interactions and then refines them with confidence-weighted ranking and cold-aware graph aggregation. Across the evaluated settings and matched recommendation backbones, ProbLLM achieves higher performance than the corresponding baselines.