At JW Player, we have developed a series of models that identify high quality ad opportunities to create a thriving video advertising marketplace. In this talk, I describe how we used signals from the video player to develop models that make several hundred million predictions daily in real time, monitored online model performance, and built new features to mitigate edge cases identified online.
At JW Player, we have developed a series of predictive ad performance models to create a thriving video advertising marketplace that empowers publishers to monetize their content and advertisers to identify high quality ad opportunities.
In this talk, I describe how we used unique signals from video plays across the web to develop models that predict user behavior and how we deployed these models to make several hundred million predictions daily in real time. I then discuss how we monitor model performance online, and share how we identified, diagnosed, and mitigated two edge cases where online model performance did not initially meet our expectations given our offline testing.