SPNet: Utilizing Subspace Projection to Achieve Feature Interaction for Click-Through Rate

Abstract

Click-through rate prediction is critical to online advertising services, wherein feature interaction is fundamental. There exist a number of schemes studying feature interaction, including factorization machine, xDeepFM, and other deep learning based models. Nevertheless, multiple features in one layer are not able to interact with each other or can not interact effectively. To resolve this problem, we propose a novel Subspace Projection Network (SPNet) in this paper. SPNet leverages subspace projection to make all features interact with each other in one subspace. Different subspaces employ different approaches to interact features in one layer. By stacking multiple layers, complex feature interactions can be implemented. To verify effectiveness of SPNet, we conduct experiments on two large-scale datasets. Experimental results demonstrate that SPNet not only outperforms the state-of-the-art shallow models, but also surpasses most deep learning based schemes.