文献类型：期刊文献

英文题名：Causal Adaptive Reweighting Meta-Learning With Dynamic Strength for Imbalanced Cold-Start Sequential Recommendation

作者：Lu, Jiahong Yu, Liya Li, Shaobo Xu, Zhao Jiang, Pei Liu, Shanhui Wang, Yingwan

第一作者：Lu, Jiahong

机构：[1] Key Laboratory of Advanced Manufacturing Technology of the Ministry of Education, Guizhou University, Guiyang, China; [2] School of Mechanical Engineering, Guizhou University, Guiyang, China; [3] School of Mechanical Engineering, Guizhou Institute of Technology, Guiyang, China; [4] School of Mechanical and Vehicle Engineering, Chongqing University, Chongqing, China; [5] Faculty of Printing, Packaging Engineering and Digital Media Technology, Xi'an University of Technology, Xi'an, China

第一机构：Key Laboratory of Advanced Manufacturing Technology of the Ministry of Education, Guizhou University, Guiyang, China

年份：2026

卷号：38

期号：14

外文期刊名：Concurrency and Computation: Practice and Experience

收录：EI(收录号：20262921132416)

语种：英文

外文关键词：Dynamics - Human computer interaction - Inverse problems - Learning systems - Population statistics - Rating - Recommender systems - Starting - Uncertainty analysis - User interfaces

摘要：Sequential recommendation systems face significant challenges in accurately modeling user preferences, particularly for cold-start users with limited interaction histories and imbalanced rating distributions. Traditional meta-learning approaches treat all samples equally, ignoring the varying informativeness of rare ratings and the heterogeneous nature of user engagement patterns. In this paper, we propose Causal Adaptive Reweighting Meta-Learning (CARML), a novel framework that integrates causal inference principles with model-agnostic meta-learning for sequential recommendation. Our approach introduces three key innovations: (1) a propensity score network that identifies and upweights rare rating samples through inverse propensity weighting, (2) a counterfactual prediction network that quantifies prediction uncertainty via entropy-based measures, and (3) an uncertainty-based dynamic causal strength mechanism that adaptively adjusts the intensity of causal correction based on user interaction frequency and prediction confidence. The dynamic strength module employs learnable combination weights to balance length-based decay factors and uncertainty-driven adjustments, enabling personalized treatment for users across the activity spectrum. Extensive experiments conducted on four real-world datasets and one industrial dataset demonstrate that CARML achieves the best RMSE across all five datasets, with an average improvement of 8.11%, and the best MAE on four out of five datasets. The most pronounced gains are observed on datasets with severe rating imbalance and limited user populations, confirming the effectiveness of causal correction under the cold-start conditions it is designed to address. ? 2026 John Wiley & Sons Ltd.