基于高低阶特征交互学习的点击率预测模型研究

Research on click-through rate prediction model based on high-low order feature interactive learning

作为在线广告推送中极为重要的环节,准确的点击率预测(Click-Through Rate,CTR)不仅能提升用户体验,更能增加经济收益,减少资源浪费。目前,基于深度学习的CTR预测模型虽然取得了一定成绩,但在高低阶特征交互学习方面存在不兼顾、不充分以及模型可解释性不强等问题。为解决上述问题,文章提出的模型基于压缩交互网络对高阶交互特征进行显式学习,增强可解释性。同时采用ECA-net网络与双线性层组合的方式,对一阶特征进行加权学习,对二阶特征进行更加细粒度的特征交互,实现深度神经网络学习更细粒度的高阶交互特征,兼顾高低阶特征学习,获取更加全面的潜在特征相关性。在Criteo和Avazu两个公开的大数据集上实验发现,与已提出的相关模型相比较,新模型在性能方面均有所提升。

As an extremely important link in online advertising push,accurate Click-Through Rate(CTR)can not only improve user experience,but also increase economic benefits and reduce resource waste.In recent years,CTR prediction models based on deep learning have achieved good results,but there are some problems in the interaction of high and low order features,such as inadequacy and model interpretability.In order to solve the above problems,the model proposed in this paper adopts the compressed interactive network to explicitly learn the high-order interaction features and enhance the interpretability.At the same time,the combination of ECA-net network and bilinear layer is used to carry out weighted learning of first-order features,carry out fine-grained feature interaction of second-order features,and realize the deep-neural network to carry out fine-grained high-order interactive feature learning,so as to obtain more comprehensive potential feature correlation.Experiments on Criteo and Avazu,two publicly available large data sets,show that compared with the existing models,the performance of the proposed models is improved.