Click-through rate prediction, which aims to predict the probability of a user clicking on an advertisement or product, is widely used and important in advertising and recommendation systems. Various click-through prediction models proposed in recent years enumerate all cross features at a predefined maximum order and then train the model to identify useful feature interactions, but there are a number of problems with this approach. First, the complexity of the model is proportional to the order, so there is a trade-off between expressiveness and computational cost. Second, at the maximum order, the introduction of some noisy crossover features can degrade the model performance. Third, implicit higher-order feature interactions are poorly interpretable and lack convincing reasons to explain the model results. In this paper, we propose MAFN, which explicitly models feature combinations of different orders using multi-headed self-attentive networks with different levels of residual connectivity. At the same time, we introduce an adaptive factorization network to learn crossover features of arbitrary order. Extensive experiment evaluations show that MAFN performs well compared to existing state-of-the-art models.
|