Abstract
For industrial-scale advertising systems, prediction of ad click-through rate(CTR) is a central problem. Ad clicks constitute a significant class of userengagements and are often used as the primary signal for the usefulness of adsto users. Additionally, in cost-per-click advertising systems where advertisersare charged per click, click rate expectations feed directly into valueestimation. Accordingly, CTR model development is a significant investment formost Internet advertising companies. Engineering for such problems requiresmany machine learning (ML) techniques suited to online learning that go wellbeyond traditional accuracy improvements, especially concerning efficiency,reproducibility, calibration, credit attribution. We present a case study ofpractical techniques deployed in Google's search ads CTR model. This paperprovides an industry case study highlighting important areas of current MLresearch and illustrating how impactful new ML methods are evaluated and madeuseful in a large-scale industrial setting.