基于深度残差网络的DeepFM点击率预测模型.docxVIP

基于深度残差网络的DeepFM点击率预测模型 李遥　李烨 摘 要：广告点击率预测是互联网广告投放系统中的核心组件，用户个性化广告推荐的准确度对于提高商业系统回报率有着至关重要的作用。提出一种基于深度残差网络的DeepFM点击率预测架构，将深度残差网络引入其中的深度神经网络，解决其随着网络变深产生的梯度消失和模型性能退化问题，从而更有效地提取高阶组合特征。为了避免过拟合引入Dropout机制，并且对输入特征的预处理引入最大支持的维度参数，以避免独热码映射时出现维度灾难。基于Criteo公开数据集的研究表明，原生DeepFM的性能优于其它主流CTR预测模型，而提出的模型则具有更低的Logloss和更高的AUC，相比原生DeepFM分别改进了1.26%和0.93%。 关键词：点击率预测;DeepFM;因子分解机;深度神经网络;深度残差网络;随机失活 DOI：10. 11907/rjdk. 192670 中图分类号：TP303 文献标识码：A 文章编号：1672-7800（2020）003-0074-05 A Click-through Rate Prediction Model of DeepFM Based on Deep Residual Network LI Yao， LI Ye （School of Optical-Electrical Computer Engineering， University of Shanghai for Science Technology， Shanghai 20XX93， China） Abstract： Advertising click-through rate prediction is a core component of Internet advertising system. The accuracy of user personalized recommendations is critical to improve the commercial income. This paper proposes a DeepFM CTR prediction model based on deep residual network. Deep residual network is introduced into the deep neural network to solve the problems of gradient disappearance and model performance degradation caused by the deepening of the network and consequently high-order feature interactions can be learned more effectively. In addition， the Dropout mechanism is adopted to prevent overfitting， and a maximum supported dimension parameter is used in the preprocessing of input features to avoid dimensional disasters. Studies on Criteo public data sets show that the native DeepFM model performs better than other mainstream CTR prediction models， while the proposed model has a lower Logloss and a higher AUC compared to native DeepFM， with improvements of 1.26% and 0.93%， respectively. Key Words： click-through rate prediction; DeepFM; factorization machine; deep neural network; deep residual network; dropout 0 引言 用戶点击率预测（Click-through Rate，CTR）研究是推荐系统中十分重要的组成模块，也是程序化广告交易框架中的智慧核心组件，学习和预测用户行为模式对个性化推荐系统、智能信息检索等领域都有着极其重要的意义。为了提升用户点击率预测准