KEML: A Knowledge Enhanced Meta-Learning Model for User Cold-Start Recommendation

Abstract

The user cold-start problem refers to the decline in recommendation quality that occurs when a new user joins a recommendation system due to sparse interaction data. To address this issue, recent studies have adopted meta-learning methods that learn global initial parameters and quickly adapt to new users with minimal interactions. Existing meta-learning recommendation methods often incorporate external data to enrich user representations, but this can introduce unwanted noise. Furthermore, they lack an effective knowledge-sharing mechanism between the different stages of meta-learning optimization. To address these limitations, we propose a novel meta-learning approach called KEML. Our method first groups similar users based on popular items to form multiple enhancement sources. A global weighted fusion strategy is then designed to generate enhanced user embeddings. By computing a distillation loss between the recommendation results from the original and enhanced embeddings, KEML refines the initialization of personalized user parameters. Additionally, the distillation loss is also applied to update the global parameters in the outer meta-learning loop, further strengthening information exchange between the inner and outer optimization stages. Experiments conducted on multiple public datasets demonstrate that KEML delivers strong performance in user cold-start scenarios. Note to Practitioners—The increasing demand for personalized content in digital platforms has highlighted a persistent challenge in recommender systems: the user cold-start problem, where insufficient historical interaction data makes it difficult to infer user preferences. Motivated by the limitations of traditional data augmentation and hybrid models—particularly their susceptibility to noise and weak adaptability—we propose KEML, a knowledge enhanced meta-learning model designed to address cold-start scenarios. KEML introduces a novel enhancement strategy by clustering similar users based on popular items and fusing their embeddings through a global weighted mechanism. A key innovation lies in our use of a distillation loss, which not only guides personalized parameter initialization in the inner loop but also informs global updates in the outer loop, thereby enhancing optimization stability and knowledge transfer. Experiments on Douban Book, MovieLens and Yelp datasets show that KEML consistently outperforms existing methods in both accuracy and ranking quality. However, the model’s performance can be affected by the quality and diversity of selected similar user groups, and the gating mechanism, while effective, may require further tuning in extremely sparse environments. Future work could explore dynamic similarity metrics, cross-domain transfer strategies, and online learning extensions to further improve robustness and scalability.