HoINT: Learning Explicit and Implicit High-order Feature Interactions for Click-through Rate Prediction

Click-through rate (CTR) prediction is a research hotspot in the field of recommendation systems and online advertising. Because of the diversity, large-scale, and high real-time characteristics of Internet data, manual feature interaction is almost impossible. Although existing models can learn feature interactions without manual feature engineering, few studies attempt to learn both explicit and implicit high-order feature interactions simultaneously. In order to effectively capture explicit and implicit high-order feature interactions, and automatically identify important feature interactions in a larger feature interaction space, we construct a parallel model that integrates a multi-head self-attention network and a Bilinear-DNN module (HoINT), which can learn high-order feature interactions automatically in both explicit and implicit ways. Sufficient experiments on four real-world datasets indicate that the HoINT model proposed is better than the most typical and advanced models, and the relative contributions of different components of the model are assessed.