A Dual‐Path Framework Based on Meta‐Learning and Attention Mechanism for Click‐Through Rate Prediction in Cold‐Start Scenarios

ABSTRACT Click‐through rate (CTR) prediction under cold‐start scenarios remains a critical challenge in recommender systems. Recent studies have primarily focused on leveraging deep learning techniques to enhance feature interaction modeling or employing transfer learning and meta‐learning strategies to address the data sparsity problem. However, existing approaches often suffer from limited feature representation capabilities and inefficient personalization when dealing with users with sparse interaction data. To address these issues, we propose a novel dual‐path framework based on meta‐learning and attention (DPMEA) for cold‐start CTR prediction. Specifically, DPMEA adopts a dual‐path parallel architecture. The global path integrates an Efficient Additive Attention mechanism into an improved AutoInt encoder, which reduces computational complexity and significantly enhances the efficiency of feature interaction modeling. The personalized path is built upon a modified Model‐Agnostic Meta‐Learning (MAML) algorithm to enable rapid user adaptation by generating personalized predictors through inner‐loop updates. An adaptive fusion layer is subsequently constructed, which leverages learnable parameters and neural networks to achieve an optimal combination of global knowledge and personalized information. Extensive experiments on the industrial cloud services dataset and the MovieLens‐1 M dataset demonstrate that DPMEA achieves substantial performance improvements, up to 27.3% better than existing methods, providing a promising approach.