实时竞价中基于展示机会细分的中标价预测

Winning Price Prediction Based on Impression Opportunity Segmentation in Real-Time Bidding

实时竞价中,准确预测展示机会的中标价能有效指导代表广告主利益的需求方平台制定竞价决策和分配广告预算,从而优化广告活动绩效。目前的主流方法是在全部样本上训练单一预测模型,忽略了不同样本之间的差异。论文提出了基于展示机会细分的中标价预测方法。根据展示机会对应的用户、媒体、广告特征,该方法首先通过聚类将展示机会的历史竞价记录细分成多个子集,针对每个子集,分别训练预测模型并进行合理组合,从而提升其预测性能。基于真实的实时竞价数据进行仿真实验,与主流方法进行对比分析,验证了该方法的有效性。

In real-time bidding,predicting the winning price of an impression opportunity accurately can effectively guide the Demand-side Platform,which represents the advertisers'benefits,to make the bidding decision and allocate advertising budget,so as to optimize the performance of advertising campaigns.The mainstream methods are to train a single prediction model on all data samples,ignoring the differences between different samples.A winning price prediction method based on impression opportunity segmentation is proposed.According to the user features,publisher features and advertising features of impression opportunities,the method divides the historical bidding records of impression opportunities into several subsets by clustering.The prediction models are trained respectively for each subset,and combined reasonably to improve the prediction performance.Simulation experiments on real-time bidding datasets and a comparative analysis with mainstream methods have verified the effectiveness of the proposed method.