一种基于逻辑回归模型的搜索广告点击率预估方法的研究

发布时间：2018-12-07 08:45

【摘要】：搜索广告是搜索引擎公司如百度,Google等的主要收入来源,当前的收费模式是按用户点击广告的次数收费,而广告位的个数是有限的,因此对于每个用户检索词,怎样在广告库中检索到用户最可能点击的广告返回给前端去展示是所有搜索引擎公司最关心的问题。CTR(Click Through Rate)即是衡量某广告被用户点击的可能性大小的指标。展示CTR最高的广告给用户是三赢的,对于用户,提高了搜索体验,他们乐于点击；对于商家,精准的广告投放策略将使广告都投放给潜在客户,有助于提高他们的营收；对于搜索引擎公司,提高用户点击广告的可能性,将使它们赚取更多广告提成。预估CTR是一项很复杂、涉及面很广的工作,本文就此问题提出了一种基于逻辑回归模型的预估方法。本文将预估CTR工作分成了两个主要方面：线下训练和线上计算。线下训练中使用Hadoop,从搜索引擎日志出发,经历数据清洗、提取特征、合并增量、排序降维、模型求解、模型验证等步骤,最终得出了一个特征到权重的映射文件,此即为我们的模型。在线上计算部分,经历扩展匹配、广告粗选、CTR计算等步骤,取出了CTR最高的10个广告返回给前端。逻辑回归模型属于有监督学习的一种,如何提取数据集中区分度高的特征是关键,本文针对CTR预估问题提出了3项一元、二元特征,经实验验证能有效提高搜索引擎收益。针对搜索广告这一特定场景,本文还引入了时间衰减因子用以区分较老的历史记录跟较新的历史记录对CTR预估的不同影响力,也取得了很好的成效。
[Abstract]:Search advertising is the main source of revenue for search engine companies such as Baidu, Google, etc. The current charging model is to charge users according to the number of times they click on advertisements, and the number of ads is limited, so for each user, How to retrieve the advertisements most likely to be clicked by a user in an ad library and return them to the front end to display is the most concerned issue for all search engine companies:. CTR (Click Through Rate) is an indicator to measure the possibility of an advertisement being clicked by a user. Showing that CTR's highest ads are triple wins for users, improving the search experience for users, they are happy to click; for merchants, accurate advertising strategy will enable ads to be delivered to potential customers, helping to increase their revenue; For search engine companies, increasing the likelihood of users clicking on ads will make them earn more ad royalties. Estimating CTR is a complex and extensive work. In this paper, a prediction method based on logical regression model is proposed. In this paper, the prediction CTR work is divided into two main aspects: offline training and on-line computing. In offline training, we use Hadoop, to start from search engine log, go through data cleaning, feature extraction, merge increment, sort dimension reduction, model solution, model verification and so on. Finally, we get a mapping file from feature to weight. This is our model. Online computing part, experienced extended matching, advertising rough selection, CTR calculation and other steps, took out the top 10 CTR ads returned to the front end. Logical regression model is a kind of supervised learning. How to extract the feature of high classification of data set is the key. In this paper, we propose three monadic, binary features for CTR prediction, which can effectively improve the revenue of search engine through experiments. In this paper, the time attenuation factor is introduced to distinguish the different influences of the older historical records and the newer historical records on the CTR estimation for the specific scene of search advertising, and good results have been achieved.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP391.3

【相似文献】