Dynamic masking-based feature interaction modeling for e-commerce click-through rate prediction

Learning feature interactions is crucial to click-through rate (CTR) prediction in e-commerce applications. The interaction between target items and user behavior history is one of the most important factors. In recent years, Deep Neural Networks (DNNs) designed to automatically learn implicit nonlinear interactions from the original sparse features have been widely adopted for industrial CTR prediction tasks. Implicit feature interactions often cannot retain the full representation capacity of the original feature interactions. Explicit modeling approaches, such as Cartesian product methods and matrix factorization-based models, either suffer from high computational costs or have limited expressive power. Recent deep learning-based methods improve feature interaction modeling but often overlook the enhancement of feature representations and fail to consider the varying importance of different interactions, limiting model performance. To address these challenges, in this paper, we report the design, implementation and deployment of Feature Dynamic Masking (FDM), a novel framework for modeling feature interactions in deep CTR prediction models. FDM dynamically generates unique feature representations for different interactions using learnable masks, enabling information sharing while maintaining efficiency. Additionally, it introduces a weighting module to capture the importance of different interactions. FDM framework is simple, effective and easy to deploy in real-world search and recommender systems. Since its deployment in the search system of Shopee Pte Ltd., Singapore, it has achieved a commercially significant 2.76% improvement in Order Per User (OPU) and 2.45% improvement in Gross Merchandise Volume (GMV).