The Science of Display – Introduction
With this blog post, I’d like to begin a series of posts on the technology and data-science that power modern display advertising (especially at Turn, where we’ve always been at the cutting-edge of innovation in this area). We will explore why the advances we’re making in the new field of computational advertising matter and how they matter. We will also try to de-mystify some of the concepts and the technologies that go into building a scalable, high performance demand side platform (DSP).
Computational advertising is an emerging field that spans multiple scientific disciplines – machine learning, optimization, statistics, information retrieval, economics, game theory, risk management, and many more. In the past, most of the work done in this field concentrated on search advertising and text ads, and not much attention was paid to display. It turns out display is a very different beast (but you already knew that). Although, ultimately, the problem is the same – how do we deliver the right message to the right person in the right situational context (not to forget, at the right price as well!)
At this point, one may wonder: what does game theory have to do with advertising? Or for that matter, how about risk management? Well, in this post, let’s examine risk with a simplified example.
With the advent of real-time bidded exchanges (RTBE), the ability to examine and bid on several billion possible ad impressions every day offers agencies and advertisers both a unique opportunity and a challenge. It allows them to cherry pick the best possible impressions (if they have sophisticated enough prediction technology), but it also exposes them to enormous risk. Simplistically speaking, even an extremely tiny error rate of 0.1% when examining 10 billion RTBE bids is 10 million wasted impressions. Without the right risk management algorithms, real-time controls, corrective feedback, and pacing algorithms, things can go south very quickly.
This is one reason why many platforms will not be able to scale up to examine every single impression available on RTBEs. Many platforms will be forced to restrict their bids to small audience segments or bid on a much smaller universe of impressions (restricted by QPS or queries per second). To truly scale performance and be able to bid effectively on a large volume of impressions, risk management science and algorithms need to be built into the platform as an integral element. At Turn, risk management is intertwined into several key components of our DSP platform.
In the next post we will examine why learning is required and how more efficient learning reduces wasted impressions (duh!).
Editor's Note: This blog post originally appeared on the Turn Blog.
Goutham Kurra is director of technology at Turn, Inc. where he focuses on developing cutting-edge technology for digital advertising including machine learning, predictive real-time bid optimization, and contextual and behavioral targeting among others. Over the last decade, Goutham has built innovative products and technology in diverse areas such as online advertising, enterprise search and e-discovery, distributed systems, computational biology, computer vision, and robotics. Prior to Turn, he was the first employee and an architect at the e-discovery pioneer Kazeon Systems (acquired by EMC) and also a founding member of the technical team at Adeosoft, where he built the world's first distributed common-language-runtime virtual computer. Goutham holds an MS degree in Computer Science from the University of Cincinnati, an MS in Physics and a BS degree in Engineering from the Birla Institute of Technology and Science, Pilani.