DRIN: Deep Recurrent Interaction Network for click-through rate prediction

Abstract

The click-through rate (CTR) prediction focuses on modelling feature interactions effectively in the domains of computational advertising and recommender systems. However, to obtain explicit feature interactions, vector-multiplication operations were usually utilized as the inner product or the outer product, which severely limited their abilities. For improving interaction performance, the Recurrent Interaction Network (RIN) was proposed, which employed two products types in the matrix-multiplication operations to fulfill the explicit feature interactions and borrowed the 1 × 1 Convolutional Neural Network (CNN) to fuse the different orders interactions. Therefore, RIN can model arbitrary order explicit interactions between two different layers, leading to increased efficiency, compared with the existing methods. Furthermore, RIN was combined with a standard Deep Neural Network (DNN) creating a new composite model, called the Deep Recurrent Interaction Network (DRIN). DRIN can learn both the implicit and explicit interactions of features, which specifically has more ability to learn explicit interactions than existing models. Comprehensive experiments show that DRIN outperforms state-of-the-art models in terms of AUC and Logloss performance.