CrossDNN: A CTR Prediction Model combining Explicit and Implicit Interactions

Abstract

CrossDNN: A simple but effective CTR prediction model for Recommendation Ying Sheng Department of Computer Software The Graduate School of Hanyang University Supervisor Professor: Inwhee Joe Click-through rate (CTR) prediction has emerged as a pivotal task in numerous practical applications. Effectively capturing intricate high-order features is crucial for ranking models. Typically, researchers employ Deep Neural Network (DNN) models to implicitly capture interactions among high-order features. Noteworthy models like xDeepFM and MaskNet have been proposed, but few attempts have been made to change the structure of the DNN model. In recent years, a substantial number of parameter adjustments have significantly enhanced the predictive accuracy of DNNs, so it is evident that DNN holds significant potential. To further unlock the potential of the DNN model, we try to change the structure of the DNN model and propose a new model, designed to enhance the DNN model's applicability in recommendation systems. Because our proposed model can make explicit feature interactions and implicit feature interactions closely intertwined, we name it CrossDNN. In the proposed model, we modified the structure of DNN by introducing several fully connected layers, so that explicit feature interaction and implicit feature interaction can be truly perfectly combined. In this paper, we make tests on two large datasets, Criteo and Avazu, and two small datasets, Movielens and Frappe. The AUC results can reach 81.46 (Criteo), 76.62 (Avazu), 97.13 (Movielens) and 98.55 (Frappe) respectively, outperforming numerous state-of-the-art counterparts. This suggests that CrossDNN exhibits effectiveness in constructing novel high-performance ranking systems.