多元逻辑回归在实时竞价中的应用研究-软件工程专业论文.docxVIP

Dissertation Submitted to Hebei University of Technology for The Master Degree of Software Engineering The Research of Multivariate Logistic Regression Algorithm on Real-Time Bidding by QiaoNing Supervisor: Prof.ShenXiTing November 2014 摘 要 实时竞价的过程是由需求方平台、媒体平台以及消费者共同完成的。这一过程 是三赢的，消费者可以看到自己感兴趣的广告，广告主可以使得投放更加精准，降 低成本，需求方平台可以赚到更多的广告提成。实时竞价是一项很复杂且涉及领域 广泛的工作，本文就实时竞价这一问题提出了一种基于逻辑回归模型的算法。将实 时竞价整体算法实现分成了两个主要方面：数据预处理和模型训练。其中模型训练 又分为线下受众选择模型训练和逻辑回归模型训练。数据处理部分，从各种竞价日 志数据入手，将其负例部分，通过一种经过优化的、对处理高维稀疏数据有很大优 势的聚类算法经行聚类，达到了很好的效果，大大减少了负例的数量。提高了正负 例比例。大大减少了后续的模型训练时长，提升了训练效果。模型训练中，从经过 处理的竞价日志数据出发，通过数据预处理、特征提取、增量合并、排序降维、模 型求解、模型验证这些步骤，得到逻辑回归模型。此模型为一个特征到权重的映射 文件。多元逻辑回归算法模型属于有监督学习的一种，关键点在于如何提取数据集 中区分度高的特征。本文针对实时竞价问题提出了二元特征，经实验结果分析，有 效的提高了总分数。并且，针对实时竞价这一特定场景，通过引入线下受众模型， 取得了很好的效果。 关键字：逻辑回归；实时竞价；聚类；稀疏特征；受众选择 ABSTRACT Real-time bidding is a process that completed by the demand side platform, the media and consumer. This process is a win-win-win, consumers can see the advertisement that they are interested, advertisers can be made into more precise, reduce cost, the demand side platform can make more advertising. Real-time bidding is a very complex and involved fields a wide range of work, in this paper, the real-time bidding proposes a algorithm based on logistic regression model. The real-time bidding integral algorithm is divided into two main aspects: data preprocessing and model training. The model training is divided into offline training audience choice model and logistic regression model. Part of data processing, log data from a variety of bidding, the negative part, by a kind of optimized, is of great advantage to deal with high-dimensional sparse data clustering algorithm in clustering, achieved good results, greatly reduce the number of the negative cases. Raise the proportion of positive and negative cases. Greatly reduces the length of training model, to improve the training effect. Model training, starting from the treated for log data, through the data preprocessing, feature ex