基于SVM回归的广告位价值预估平台设计与实现

发布时间：2018-04-16 09:37

本文选题：广告 + 回归模型　；参考：《南京大学》2012年硕士论文

【摘要】：广告是众多互联网站点的主要盈利模式之一,随着互联网广告行业的发展,有越来越多的网站将广告位交由广告联盟托管。广告联盟和多个广告主签订广告投放计划,通过一系列匹配算法,将广告展现在最合适的网站上。广告普遍按照点击次数计费,但广告的点击率是不确定的,不同广告的点击单价也不同。对于网站主来说,一个广告位未来能获取的收益是未知的,这在一定程度上影响了网站主的积极性。在这种情况下,如何预测一个广告位所能获取的收入就成为一个亟待解决的问题。本文介绍了一种预测广告位收益的方法,主要思想是利用大量历史数据,通过支持向量机来建立一个回归模型。主要工作分为三个部分。第一,环境特征抽取,环境特征通过抓取并分析广告位所在网页的内容得到。第二,数据处理,分析所有特征数据的有效性和特征之间的相关性,之后根据模型训练的需要,对数据进行筛选和转换。第三,模型训练与优化,模型训练基于已有的支持向量机算法库完成,在此基础上通过参数寻优、特征选取、特征绑定等方法做了大量优化。本文详细阐述了这三个阶段的工作和成果,并在最后进行了简单总结,提出了未来可以进行的改进。
[Abstract]:Advertising is one of the main profit models for many Internet sites. With the development of the Internet advertising industry, more and more websites are handing over advertising positions to ad federations.The Advertising Alliance has signed up with multiple advertisers to display ads on the most appropriate site through a series of matching algorithms.Ads generally charge according to the number of clicks, but the click rate is uncertain, and the click price varies from ad to ad.For website owners, the future revenue of an advertising site is unknown, which to some extent affects the enthusiasm of site owners.In this case, how to predict the revenue a advertising position can achieve becomes a problem to be solved.This paper introduces a method to predict the revenue of advertising bit. The main idea is to build a regression model by using a large amount of historical data and support vector machine.The main work is divided into three parts.First, environmental feature extraction, which is obtained by grabbing and analyzing the content of the page in which the advertisement is located.Secondly, data processing analyzes the validity of all feature data and the correlation between features, and then filters and transforms the data according to the needs of model training.Third, model training and optimization, model training based on the existing support vector machine algorithm library, on this basis through parameter optimization, feature selection, feature binding and other methods to do a lot of optimization.In this paper, the work and achievements of these three stages are described in detail, and a brief summary is made at the end of this paper, and some possible improvements in the future are put forward.
【学位授予单位】：南京大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP393.092

【参考文献】