实时广告竞拍平台中的海量数据分析和竞价预测  

Mass data analysis and bid price forecasting in online ad exchange marketplace

作　　者：毛衡[1] 胡宁 陈蔚[1] 高卫国[1] 陈文斌[1] 

机构地区：[1]复旦大学数学科学学院,上海200433 [2]聚越信息技术(上海)有限公司,上海200070

出　　处：《应用数学与计算数学学报》2016年第1期1-15,共15页Communication on Applied Mathematics and Computation

基　　金：国家基础科学人才培养基金资助项目(J1103105)

摘　　要：实时广告竞拍平台会产生海量的数据,如何对这些数据进行分析和建模会决定广告竞拍的成败.其中一个重要的问题是,如何得到胜出竞价的概率密度函数以便用于指导竞价.在文献(Cui Y,Zhang R,Li W,Mao J.Bid landscape forecasting in online ad exchange marketplace.Proceedings of the 17th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining.New York:ACM,2011:265-273)中,雅虎实验室提出了一个技术框架来解决这个问题:首先利用合理的统计学方法对海量数据按照特征属性进行分类;再利用高效的数据结构进行存储,以便快速定位特定属性的信息;最后用梯度提升决策树(gradient boosting decision trees,GBDT)模型和有限混合模型(finite mixture model,FMM)学习得到胜出竞价的分布模型.结合国内竞拍平台的数据特点,在上述文献的基础上,对技术框架进行改进,并提出修正的算法.Online ad exchange marketplace will produce huge amounts of data. The analyses and modeling of the data will determine the results of advertisement action. One of the important problems is how to derive the cumulative probability density function of win price. In the reference （Cui Y, Zhang R, Li W, Mao J. Bid landscape forecasting in online ad exchange marketplace. Proceedings of the 17th A CM SIGKDD International Conference on Knowledge Discovery and Data Mining. New York： ACM, 2011： 265-273）, Yahoo! Lab proposed a technical framework to overcome this problem： Firstly, features and attributes of mass data are classi- fied by rational statistical methods and stored by efficient data structure. Then, the information of specified attributes can be located quickly. Finally, the distri- bution of the win price is obtained by the GBDT （gradient boosting decision trees） model and the FMM （finite mixture model）. In the paper, we use the property of the domestic ad exchange marketplace to modify the technical framework and propose a modified algorithm based on the framework of the reference.

关 键 词：广告交易市场 竞拍价预测 带星树 

分 类 号：O212.1[理学—概率论与数理统计]