Modeling Advertiser Bidding Behaviors in Google Sponsored Search with a Mirror Attention Mechanism
Suqi Liu, Liang Liu, Sugato Basu and Jean-François Crespo
In this paper, we present a data-driven approach to predicting advertiser’s bid prices in a sponsored search system, like Google SearchAds, with a new type of attention-based sequence learning model.Instead of characterizing the advertisers’ bidding behaviors with explicit assumptions (e.g., rationality models), as done in previous work , we treat their bid adjustments as response to observable metrics (e.g., impression count, click-through rate) and directly predict the bid prices using recurrent neural networks combined with a novel attention mechanism. The proposed model consists of two recurrent neural networks, for capturing the dynamics of metric sequence and bid sequence respectively, connected by a mirror attention layer formulation that transfers location information from metrics to bids. We evaluate the performance of the proposed model, along with other baselines, on advertiser bidding history data extracted from Google Search Ads system logs. We also demonstrate the generality of the new mechanism by experimenting on another domain: air quality prediction. Our empirical results show the effectiveness of the modeling approach and the new mechanism— we see a significant improvement over the baseline models for both advertiser response prediction and air quality prediction tasks.