Dr. Boris Mouzykantskii, IPONWEB Delivers Keynote Speech at ATS London 2016

Thanks for your kind words. I'll speak about machine learning and how it's used in marketing. Two slides about our company: we're a technology company that has helped many ad tech companies build media bid systems. We offer a framework for custom solutions and, more recently, Bit Switch to plug into the RTB ecosystem quickly. Why machine learning? It's the new big data — Hadoop and big data terms have peaked, but ML is rising. Agencies and advertisers will need to enable ML in their big data systems. Crazy money will be channeled through these systems from new entrants, not just traditional ad tech players. Conceptually, you take a business process (like matching ads to users), it generates data, you feed that into an ML system, and whatever it learns feeds back. The key is that humans are not in the iterative loop; they sit aside and tweak. This helps because machines start wider, iterate faster, and discover more value. For years we said building your own bidder was the way to go, but now we think you need a technology or DSP partner to handle the complexity. From an architecture perspective, you detach the ML brain from the monolithic bidder. Let me explain RTB 101: an opportunity is a combination of an SSP bid and a creative. It has data and a predicted value to the advertiser. To win, you place a bid and pay a price that depends on the bid. The key is that value, bid, and price are all different. To split the ML system from the execution platform: the advertiser supplies a function that evaluates value given data and determines daily spend. Because they supply the whole function at once, they can push updates infrequently. The execution platform must maximize predicted value under a daily budget. The optimal solution is a straight line on the opportunity cost, adjusted by slope. The tricky part is learning the bid-to-win. On the ML side, they compute conversion probability from logs: add conversion indicator, feed into ML tool, output a decision tree or regression. They also figure out daily spend by plotting value vs. cost. That's the theory. We run about 6,000 models a day. What can go wrong? First, the curse of dimensionality — effective dimensionality between a million and a billion. Second, data collection bias: you only learn on data you collected, so your world is limited. The runner will break out of that boundary, making your dataset incomplete and biased. You need to live with that. Third, the communication gap between data science and business, like the Cat and Alice. Data scientists can find spurious correlations (e.g., Internet Explorer causing deaths, organic food causing autism). Business must help distinguish correlation from causation. Fourth, supply side auctions aren't always simple second price; they use mixed auctions or cheat. The runner needs ML to compensate. Once you solve all that, you have a closed-loop ML system. Machines are totally committed and lack morality. They do exactly what you tell them, and we are bad at explaining what we want. Example: eBay vs. Sean Hogan. eBay wanted traffic and sales, but the affiliate company was paid for getting a cookie at purchase. They discovered cookie stuffing was more efficient, costing eBay $28 million and the guy went to jail. If you replace this with a closed-loop ML system, your engine will quickly discover that cookie stuffing is more efficient — in days, not months. It won't go to jail; it might be seen as a huge success until caught. Conclusion: ML is coming, you need a trusted partner, learn ML properly, and after doing everything perfectly, watch carefully what your system is doing. Good luck.