A multimodal approach for improving market price estimation in online advertising

Abstract

Learning the distribution of market prices is an important and challenging issue for demand-side platforms (DSPs) that serve as advertisers’ agents to compete for online advertising placements in real-time bidding (RTB) systems. Many existing approaches make an assumption that the market prices follow an unimodal distribution. However, based on analytical insights from real-world datasets, we found the distinct multimodal characteristics underlying the distribution of market prices. Moreover, the impression-level features for each ad are also ignored by these approaches in prediction, reducing the accuracy further. To address these problems, a Gaussian Mixture Model (GMM) is proposed in this paper to describe and discriminate the multimodal distribution of market price by utilizing the impression-level features. To further improve its robustness, GMM is extended into a censored version (CGMM) to handle the right-censored challenge in RTB systems (i.e., the market price is only visible to the winner of the ad auction. Thus, the dataset is always biased). Extensive experiments on two real-world public datasets demonstrate that GMM and CGMM significantly outperform 10 state-of-the-art baselines in terms of Wasserstein distance, KL-divergence, ANLP and MSE. To the best of our knowledge, this paper is the first work to simultaneously deal with the multimodal nature of market price distribution and the right-censored challenge in existing RTB systems. It will enable future RTB systems to develop more realistic bidding strategies to enhance the efficiency of online advertising placement auctioning.