SEM: A Softmax-based Ensemble Model for CTR estimation in Real-Time Bidding advertising
SEM: A Softmax-based Ensemble Model for CTR estimation in Real-Time Bidding advertising
复制标题
DOI：
10.1109/bigcomp.2017.7881698
复制
发表时间：
2017
期刊：
影响因子：
--
通讯作者：
Wen-Yuan Zhu;Chun-Hao Wang;Wen-Yueh Shih;Wen-Chih Peng-;Jiun-Long Huang
中科院分区：
文献类型：
--
作者：
Wen-Yuan Zhu;Chun-Hao Wang;Wen-Yueh Shih;Wen-Chih Peng-;Jiun-Long Huang
在实时竞价广告中，评估竞价请求和广告的点击率对于需求方平台上的竞价策略优化具有重要意义。基于回归的方法是RTB中CTR估计的流行方法，因为这种方法是高效和可扩展的。投标请求和广告的信息包含分类属性（例如URL）和数字属性（例如广告大小）。为了向量化基于回归的方法的输入信息，分类属性一般将扩展为几个二进制特征。然而，某些分类属性有无限可能的值（如URL）。因此，对于这些属性，只有训练中的观测值将被转换为二进制特征。如果在线环境中有新的属性或值，则矢量化后这些信息将丢失。在本文中，我们首先利用特征散列技巧将分类和数值属性转换为大的固定大小的向量。由于向量是大而稀疏的，我们提出了一个基于Softmax的Ensemble模型，SEM，它只采用了几个关键特征后，特征散列CTR估计。实验结果表明，该方法能够适应RTB中的恶劣环境，并且在两个真实的数据集上采用少于50个特征的情况下，有效地优于现有方法。
In Real-Time Bidding (RTB) advertising, evaluating the Click-Through Rate (CTR) of a bid request and an ad is important for bidding strategy optimization on Demand-Side Platforms (DSPs). The regression-based approaches are popular for CTR estimation in RTB since this kind of approach is highly efficient and scalable. The information of the bid request and the ad contains categorical attributes (such URL) and numerical attributes (such ad size). To vectorize the information for the input of regression-based approaches, the categorical attributes will be expanded to several binary features in general. However, some categorical attributes have infinite possible values (such as URL). Thus, for these attributes, only observed values in training will be transformed into binary features. If there is a new attribute or value in online environment, this information will be lost after vectorization. In this paper, we first exploit the feature hashing trick to transform the categorical and numerical attributes into the large fixed size vector. Since the vector is large and sparse, we propose a Softmax-based Ensemble Model, SEM, which adopts only a few key features after feature hashing for CTR estimation. The experimental results demonstrate that our proposed approach is able to adapt to the harsh environments in RTB, and outperforms the state-of-the-art approaches effectively when only less than 50 features are adopted in two real datasets.