A gate-enhanced neuro additive graph neural network via knowledge distillation for CTR prediction

Abstract

The click-through rate (CTR) prediction is a crucial task in commercial recommender systems and online advertising platforms. Recent studies have revealed shortcomings in CTR optimization, particularly in their ability to effectively identify and interpret the abnormal or latent feature interactions obscured by user behaviors. In this paper, a novel CTR prediction model is developed in two stages. The first stage formulates a gate-enhanced neuro additive graph neural network (GNAGNN), by dynamically capturing the complex interactions between features in different input environments, the adaptability of the model to the importance of features is significantly improved. While the second stage utilizes the knowledge distillation framework, enabling GNAGNN to effectively learn from a gated ensemble of existing CTR models. Unlike most higher-order feature interaction models that rely on deep neural networks, our method avoids high-complexity matrix computation and significantly reduces the computational overhead. Specifically, the framework adopts a dynamic parametric mechanism to determine the weight of the model involved in the prediction through the continuous action vector, and then achieve the accurate prediction of each advertisement impression. Eventually, comprehensive experiments carried out on two public datasets convincingly demonstrate that GNAGNN significantly outperforms the state-of-the-art baselines, and it can offer precise and interpretable understandings into features and their interactions while reducing computational costs.