Rec2: Embedding Table Reconstruction for Deep Recommender Systems

Recommender systems are a pivotal component of modern Internet services. Among them, embedding tables play a crucial role in predicting user preferences and behaviors. However, existing studies mainly focus on improving downstream recommender models, while neglecting the structural importance of the embedding table itself. As a result, they fail to fully exploit the potential of feature fields and their interactions. To address this limitation, we propose the Rec2 framework (Reconstruction for Recommender Systems), which aims to optimize the organization of embedding tables by efficiently determining which features should be split or aggregated. We introduce a novel method that estimates feature importance based on second-order Taylor expansion contributions, enabling precise identification of features that benefit from splitting or pre-constructed interactions. Moreover, we further integrates second-order estimation with a global heap mechanism to efficiently manage interactions among key features and their aggregations. By improving the structure of the embedding table, Rec2 significantly enhances recommender system performance and shows great potential for advancing personalized recommendation. The code is available at https://github.com/xqwustc/Rec2.