Real-Time Bid Prediction using Thompson Sampling-Based Expert Selection

We study online meta-learners for real-time bid prediction that predict by selecting a single best predictor among several subordinate prediction algorithms, here called "experts". These predictors belong to the family of context-dependent past performance estimators that make a prediction only when the instance to be predicted falls within their areas of expertise. Within the advertising ecosystem, it is very common for the contextual information to be incomplete, hence, it is natural for some of the experts to abstain from making predictions on some of the instances. Experts' areas of expertise can overlap, which makes their predictions less suitable for merging; as such, they lend themselves better to the problem of best expert selection. In addition, their performance varies over time, which gives the expert selection problem a non-stochastic, adversarial flavor. In this paper we propose to use probability sampling (via Thompson Sampling) as a meta-learning algorithm that samples from the pool of experts for the purpose of bid prediction. We show performance results from the comparison of our approach to multiple state-of-the-art algorithms using exploration scavenging on a log file of over 300 million ad impressions, as well as comparison to a baseline rule-based model using production traffic from a leading DSP platform.