实时竞价系统中出价算法的研究与实现

发布时间：2018-06-20 17:46

本文选题：实时竞价 + 点击率预测　；参考：《电子科技大学》2017年硕士论文

【摘要】：随着互联网技术的快速发展及广泛普及,越来越多的广告活动逐渐从线下转换到线上。目前,互联网广告中的展示广告的一种主要投放渠道是通过实时竞价系统(Real Time Biding,RTB)投放。在RTB中,需求方平台需要对每个到来的广告曝光机会,按照广告活动投放规则计算其竞标价格,竞标价格将直接决定该广告是否能被投放,因此是RTB中的一个关键问题。作为一种新颖的广告投放方式,RTB中的许多关键算法还存在较大的优化空间,本论文将针对面向展示广告的实时竞价系统,特别是其中的出价问题展开研究。本论文首先针对预算受限下的出价问题展开研究,将其建模为一个预算约束受限下的收益最优化问题,然后分别针对其中的点击率预测模型以及预算管理框架下的出价函数两方面展开研究。在点击率预测模型方面,本论文提出一个基于张量和深度学习的点击率预测模型,该模型首先将点击率预测涉及到的三个方面的特征建立为三阶张量,通过笛卡尔乘积得到三元特征组合;然后利用多项式回归方法学习每个特征的隐含向量和权重;最后利用学习得到的隐含向量和单个特征的权重作为深度神经网络的输入,重新基于训练集学习点击率预测模型。在RTB的出价算法和预算管理方面,本论文设计了一个基于历史成交价格和失败竞标价格的中标率预测模型,从而有效提升了需求方平台出价算法的有效性。此外考虑到每个广告活动在其投放周期内的预算有限,为了使得广告活动在每个时段都能平滑地投放,本论文还设计了一个基于时段广告流量和中标率的广告预算的预分配和实时调整方案。点击率预测模型在iPinYou的数据集上对提出的点击率预测模型进行了性能评估,并将其与目前主要的预测模型的性能进行了对比分析,结果表明本文提出的预测模型的AUC指标比因子分解机的预测提升了2.7%,比逻辑回归预测提升了3.8%。而基于iPinYou的大规模实验表明,按照本文提出的中标率预测模型的出价算法,其获得的用户点击数比提升了2.8%。在本文提出的预算管理框架下,用户点击数比不采用预算管理提升了8%。
[Abstract]:With the rapid development and widespread popularity of Internet technology, more and more advertising activities are gradually converted from line to line. At present, the main channel of display advertising in Internet advertising is to put in the Real Time Biding (RTB). In RTB, the demand side platform needs every coming advertisement exposure machine. It is a key problem in RTB. As a new way of advertising, many key algorithms in RTB still have larger optimization space. This paper will be aimed at real time bidding for display advertising. This paper first studies the bidding problem under budget constraints, and model it into a revenue optimization problem with limited budget constraints, and then studies the two aspects of the click rate forecasting model and the price function under the framework of budget management. In this paper, a clicking rate prediction model based on tensor and depth learning is proposed in this paper. First, the three characteristics of click rate prediction are set up as three order tensors, and the combination of three element features is obtained by Descartes product. Then the implicit vector of each feature is learned by multi term regression method. Finally, using the implicit vector of learning and the weight of a single feature as the input of the deep neural network, we re based on the learning click rate prediction model of the training set. In the RTB bidding algorithm and budget management, this paper designs a prediction model based on the rate of history transaction price and the failure bid price. It effectively improves the effectiveness of the demand side platform bid algorithm. In addition, considering the limited budget of each advertising campaign in its launch cycle, in order to make the advertising activities put smoothly at every time period, this paper also designs a pre allocation and real-time adjustment scheme based on the traffic and the rate of bid rate. The clicking rate prediction model performs a performance evaluation on the proposed click rate prediction model on the iPinYou data set, and compares it with the performance of the main prediction model. The results show that the prediction model proposed in this paper has been improved by 2.7% compared with the factor decomposer, which is better than the logical regression prediction by 3.8%.. And that large-scale experiments based on iPinYou, according to the bidding algorithm proposed in this paper the successful rate prediction model, the user clicks than to enhance the budget management framework 2.8%. is proposed in this paper, the user clicks the 8%. is raised by not using budget management
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：F713.8;TP18

【相似文献】