展示广告中点击率预估问题研究

发布时间：2018-10-10 11:02

【摘要】：随着互联网技术的发展以及随之而来的信息高流动性,互联网广告成为了商家推崇的一种主流营销方式,广告收入也已经成为互联网公司收入的重要组成部分。广告点击率(Click Through Rate,简称CTR)预估在精准广告投放过程中扮演了很重要的角色,预估的准确性对广告主的收益、广告商的收益以及用户的友好体验有着重大的影响,因此受到互联网企业的广泛关注。在本文中,我们重点关注展示广告(Display Advertising),系统地介绍和分析了在线广告系统的组织结构以及参与对象,阐述了广告点击率预估在广告系统中的重要地位。本文重点关注广告系统中点击率预估三个方面的问题。第一个是统一特征平台的构建。考虑到数据有多个不同的来源,数据内容也包含多个组成部分,如何从原始数据中提取出有用的特征并高效地将这些信息进行整合供算法使用有很大的改进空间。本文提出了在真实应用场景中体系化地构建特征、做好特征工程工作的方法,可以从不同的原始日志信息中提取出有用的特征,构建出相对干净的数据特征集合。第二个是高效点击率预估模型的提出。现有的工作已有很多将机器学习算法应用到点击率预估中,但是现有的模型多以线性模型为主,无法建模出广告信息与用户信息间的关联关系,模型的改进有很大的空间。本文提出了对偶群稀疏模型,构建出广告系统参与对象间的关联关系,从而改进点击率预估的准确性,同时能够在所有特征中做一个特征选择,以促进高效的特征工程工作以及快速的线上预测工作。第三个是大规模应用场景下的分布式算法实现以及应用。现实应用场景中,存在着数据量大、计算量大的问题,本文提出了基于MPI(Message Passing Interface)的算法分布式实现,使得模型能够充分利用计算集群资源去从海量的数据中学习出准确的模型,从而在真实场景中得到应用。
[Abstract]:With the development of Internet technology and the high mobility of information, Internet advertising has become a mainstream marketing method, and advertising revenue has become an important part of the revenue of Internet companies. Ad click rate (Click Through Rate,) prediction plays a very important role in the process of accurate advertising. The accuracy of the prediction has a significant impact on the advertisers' income, advertisers' earnings and the friendly experience of the users. Therefore receives the Internet enterprise's widespread concern. In this paper, we focus on the introduction and analysis of the organizational structure and the participating objects of the online advertising system based on (Display Advertising), and the important position of the ad click rate prediction in the advertising system. This paper focuses on three aspects of the prediction of click rate in advertising system. The first is the construction of unified feature platform. Considering that the data has many different sources and the data content also contains many components, how to extract the useful features from the raw data and efficiently integrate the information for the use of the algorithm has great room for improvement. In this paper, a method of systematically constructing features in real application scenarios and doing well in feature engineering is proposed, which can extract useful features from different original log information and construct relatively clean data feature sets. The second is the high-efficiency click rate prediction model. A lot of existing work has applied machine learning algorithm to the prediction of click rate, but most of the existing models are linear models, which can not model the relationship between advertising information and user information, so there is a lot of room for improvement of the model. In this paper, a sparse dual group model is proposed to construct the correlation relationship between the objects involved in the advertising system, so as to improve the accuracy of the prediction of the click rate, and at the same time to make a feature selection among all the features. In order to promote efficient feature engineering and fast online prediction work. The third is the implementation and application of distributed algorithms in large scale application scenarios. In the practical application scene, there are many problems, such as large amount of data and large amount of computation. This paper proposes a distributed algorithm based on MPI (Message Passing Interface), which makes the model make full use of the computing cluster resources to learn the exact model from the massive data. In order to be used in the real scene.
【学位授予单位】：上海交通大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：F713.8

【相似文献】