Abstract

Collaborative Filtering (CF) is effective for personalized recommendation but faces critical challenges with cold-start items lacking interaction data. Existing content-based and distillation approaches struggle to capture complex semantic-to-collaborative mappings and produce suboptimal embeddings due to deterministic constraints. To address this, we propose CGDRec (Content-Guided Diffusion model for Recommendation), which generates high-quality CF embeddings through conditional diffusion. Our hybrid backbone combines Mixture-of-Experts for adaptive content modeling and Cross-attention for aggregating collaborative signals from similar warm items. Experiments on three dataset ̶CiteULike, XING, and ML25M show CGDRec consistently outperforms baselines, significantly improving cold-start recommendation quality. This is the first application of conditional diffusion to cold-start item recommendation.