A Review of Deep Learning and Large Language Models for Cold Start Problem in Recommender Systems

ABSTRACT Recommender systems are essential for information filtering but often suffer from the cold start problem caused by limited interaction data. Recent advances in deep learning (DL) and large language models (LLMs) have shown promise, yet systematic analysis of their effectiveness remains scarce. To address this gap, we introduce a paradigm‐driven taxonomy that categorizes solutions by their primary source of information: content, structure, transfer, and generation. Within this framework, DL methods have matured in leveraging content and structural information from interaction logs and multimodal data, while LLMs demonstrate advantages in text‐rich and data‐sparse environments through transfer‐based paradigms that exploit semantic understanding and pre‐trained knowledge. Furthermore, emerging generative approaches show potential for synthesizing data or relations to alleviate information scarcity. No universal solution exists; effectiveness depends on the dominant paradigm of a given scenario as well as data availability and computational cost. Combining DL and LLM offers substantial opportunities, including enhanced feature representation, data augmentation, and hybrid pipelines. However, research gaps persist, particularly the lack of standardized evaluation metrics and limited exploration of integration strategies. Addressing these challenges through a paradigm‐aware perspective could significantly improve the robustness and adaptability of the cold‐start recommendation in diverse contexts. This article is categorized under: Application Areas > Data Mining Software Tools Technologies > Machine Learning Technologies > Artificial Intelligence