CTR Prediction Model Considering the Importance of Embedding Vector

Click-through rate (CTR) prediction plays an important role in industrial bidding advertising. Most of the deep CTR models focus on the interaction problem of capturing features after embedding layer, such as PNN, NFM, DeepFM and xDeepFM. Few models consider the importance of features before feature interaction. The ECA module proposed in CVPR2020 computer vision is an extremely lightweight plug and play module, which is used to improve the performance of various deep CNN architectures. This paper introduces the idea of ECANET to modify the model embedding layer to dynamically learn the importance of embedding features and construct a new deep CTR prediction model. ECANET and NFM model are combined to build a novel model ECANFM. The effectiveness of the model is verified by a comparative experiment on a public data set.