Greatmeta: gradient-aware adaptive meta-learning for cold-start recommendations

• Published in: Data mining and knowledge discovery Vol. 39; no. 5; p. 42
• Main Authors: Du, Yantong, Chen, Rui, Han, Qilong, Tan, Qiaoyu, Song, Hongtao, Zhang, Chi
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.09.2025; Springer Nature B.V
• Subjects: Adaptation; Artificial Intelligence; Chemistry and Earth Sciences; Cold; Computer Science; Data Mining and Knowledge Discovery; Design factors; Information Storage and Retrieval; Learning; Physics; Preferences; Recommender systems; Statistics for Engineering; User imbalance; Recommender systems; Cold start; Meta learning
• ISSN: 1384-5810; 1573-756X
• DOI: 10.1007/s10618-025-01123-5

In recent years, optimizer-based meta-learning, i.e., model-agnostic meta-learning (MAML), has emerged as a powerful tool to tackle the cold-start problem in recommender systems. However, existing methods overlook the key fact that user preferences are inherently imbalanced in real-world applications. Learning indiscriminately from imbalanced user preferences may lead to an imbalanced meta-model that introduces systematic bias in subsequent recommendations. In this paper, we explore the impact of imbalanced user preferences on the meta-training process and consequently propose a novel Gradient-aware adaptive Meta-learning (GreatMeta) model for cold-start recommendations. Inspired by prior work, our key idea is to leverage gradient signals to understand a meta-model’s status and guide the learning process. More specifically, we put forward three original gradient-aware factors to measure the state of the meta-model and the level of user imbalance so that we can make use of the correlation between the meta-model and the underlying users to generate a balanced meta-model. Based on these factors, we design a preference-aware scheduler to adaptively adjust each user’s contribution in the meta-training process, which helps the resultant meta-model better generalize to minority cold-start users. We also introduce a personalized encoder that effectively utilizes limited data and accelerates the adaptation process of the balanced meta-model. Moreover, we theoretically justify the rationality of the three proposed gradient-aware factors. Extensive experimental results on public benchmark datasets demonstrate the superiority of GreatMeta over a large number of state-of-the-art recommendation methods, confirming the value of addressing the imbalance of user preferences in MAML-based meta-learning for cold-start recommendations. The code is available at https://github.com/YantongDU/GreatMeta .