Dynamic Explicit Embedding Representation for Numerical Features in Deep CTR Prediction
Yuan Cheng · Proceedings of the 31st ACM International Conference on Information & Knowledge Management · 2022
Click-Through Rate (CTR) prediction is a key problem in web search, recommendation systems, and online advertising display. Deep CTR models have achieved good performance due to adoption of the feature embedding and interaction. However, most research has focused on learning better feature interactions, with little attention to embedding representation. In this work, we propose a Dynamic Explicit Embedding Representation (DEER) for numerical features in deep CTR prediction, which can provide explicit and dynamic embedding representation for numerical features. The DEER framework is able to discretize numerical features automatically and dynamically, which can overcome the discontinuity problem in the representation of numeric information. Our methods are tested on two public datasets, and the experimental results show DEER can be applied to various deep CTR models, which also improve the performance effectively.