Interpretable click-through rate prediction through distillation of the neural additive factorization model

Abstract

An accurate estimation of the click-through rate (CTR), that is, the probability of clicking on a recommended advertisement item online, is crucial for advertising agencies to make appropriate suggestions to users and maximise their income. However, the existing CTR prediction techniques exhibit two major limitations in estimating accurate CTRs. First, they lack interpretability as they are unable to justify and explain the outcomes of the models. Irrelevant recommendations of products and services may have severe consequences on the personal and financial safety of a user. Second, the high computational complexity of deep-learning models hinders deployment in real-time recommender systems. Thus, this study proposes an interpretable, accurate, and efficient CTR estimator based on the neural additive factorization model (NAFM). The NAFM leverages the model distillation framework, where it effectively learns from a gated ensemble of existing CTR models. The NAFM provides interpretable insights into features and their interactions with a reduced computational cost. The effectiveness and efficiency of the proposed NAFM was verified by applying it on two public click-through datasets, namely, Criteo and Avazu. Unlike post-hoc explainable models, the NAFM avoids an additional phase of training to generate an interpretable model, and its explanations are truthful to the model.