Wide & ResNet
Xuxu Gao, Hongxia Bie · 2018
The model Wide & Deep is widely applied for CTR prediction in recommender systems, which unites linear model for memorization and deep neural network for generalization. However, the essence of Wide & Deep still lies in feature engineering, especially in the processing of categorical features, which usually requires manually crafted feature engineering to generate all types of cross feature. Some explicit and easy-to-understand feature interactions can be extracted manually, but more feature interactions are hidden. In this paper, we propose an improved network structure for CTR Prediction, Wide & ResNet (WRN), which still keeps the linear model Logistic Regression but introduce the idea of residual network in DNN component. The advantage of doing this is that not only enhance features reuse but also learn the interactions between the low-degree features of shallow layers and the highly nonlinearity features of deep layers, so that more hidden feature interactions can be mined. Experimental results on two real-world dataset: Criteo and Frappe dataset, show that Wide & ResNet significantly outperforms the Wide & Deep model.