【作者】 徐田；

【导师】 刘秉权； 牛艳桃；

【作者基本信息】 哈尔滨工业大学 ， 软件工程（专业学位）， 2017， 硕士

【摘要】 随着互联网产业的成熟以及用户规模的扩张,互联网广告的营销价值也随之不断攀升。若能利用互联网广告的天然优势,通过点击率预估技术正确追踪用户对某广告的偏好,可带来广告主转化提升、用户体验提升、发布者收入提升等多维度的收益。因此,本文选择对计算广告学中的展示广告点击率预估问题展开研究,并将在研究过程中搭建展示广告点击率预估平台,为当前问题提供完整的机器学习解决方案。本文将系统地介绍展示广告点击率预估平台的构建过程。首先通过对点击率预估问题研究意义及研究现状的分析,引出本文的研究内容。随后,结合点击率预估问题的机器学习方案步骤对点击率预估问题展开业务分析,明确点击率预估平台的功能与性能需求,即针对计算广告中的海量数据,支持多种模型的独立、混合使用,配备离线批量学习和在线学习两种训练模式,为用户提供了从特征工程、模型训练、模型评估、模型预测到结果分析的一站式服务。紧接着,围绕需求分析的结果开始阐述点击率预估平台的详细设计及实现过程,其中特征工程与模型训练的设计与实现是文本的研究重点。由于点击率预估问题的数据来源往往是线上真实的服务日志,本文将通过系统的特征工程挖掘湮没在大量噪声中的有效特征,并力求使用最少的特征带来最佳的模型预测效果。模型训练阶段,本文选择了适合离散高维特征场景的逻辑回归模型,以及适合稀疏特征场景的因子分解机模型,并将其分别与上游的GBDT模型通过Stacking集成算法进行融合,达到提升模型预估效果的目的。平台初步实现后,将通过功能测试与性能测试,发现平台存在的问题。通过进一步的优化与迭代,完成平台的全部搭建工作。最后,对论文内容进行了总结,并对平台的改进方向进行了展望。更多还原

【Abstract】 As the development of the Internet industry and the increase of the Internet users,the marketing value of Internet advertising will continue to ascent.If we can grab the natural advantages of Internet advertising,try to use the click-through rate prediction to track the user’s preference for an advertisement,it will bring multiple benefits such as impoving the profit of advertisers,the income of publisher and the users experience.Therefore,this paper chooses to study the click-through rate prediction of display advertising and a platform is builded finally,which try to provide a complete machine learning solution for the the click-through rate prediction.This paper systematically introduces the design and implementation of the clickthrough rate prediction platform for display advertising.Firstly,the paper shows its research content after analysing the significance and the present development of the click-through rate prediction.Subsequently,analysing the requirement of the platform according to the machine learning solution steps,that is,the platform provides users with the feature engineering,the model training,the model evaluation,the model prediction and the results analysis.Then,the design and implementation of the platform are discussed around the requirement.In this part,the paper focus on the feature engineering and the model training.The platform excavates the effective features in the real service log through the feature engineering,and try to use the least feature to bring the best model prediction.After receiving the output of the feature engineering,the model training begins working.I choose the logistic regression model which is suitable for processing the high-dimensional features,and the factorization machine model which is suitable for processing the more complex features.Meanwhile,these models are integrated with the gradient boosting decision tree model to improve the prediction result.After the initial implementation of the platform,the paper start to test its quality and try to fix all the problems through further optimization and iteration.Finally,it summarizes the content of this paper and looks forward to the improvement direction of the platform.更多还原