Accurate and interpretable CTR prediction via distilled neural additive feature interaction network

Click-through rate (CTR) prediction, which estimates the probability that a user will click on an advertisement or an item, is critical to online advertising recommender systems. A key factor in optimizing CTR is understanding how to discover and explain uncommon or hidden feature interactions concealed behind user behaviors. In this paper, a neural additive feature interaction network model is firstly constructed (abbreviated as NAFI), which can automatically learn the low- and high- order feature interactions of input features with good explainability. Then a multi-teacher knowledge distillation network is utilized to realize the lightweight of NAFI (abbreviated as KD-NAFI). Finally, comprehensive experiments on three public datasets demonstrate the accuracy and interpretability of our models.