In this talk I will present new approaches for experimentation and optimization in complex software systems. A canonical example is a personalization engine for ranking search results, where decisions are combinatorial and must be made on the basis of rich contextual information like user features. These problems involve partial feedback — since we cannot measure how a user interacts with search results that were not displayed — which presents fundamental challenges for learning and optimization.
In the first part of the talk, I will describe a new online optimization algorithm that overcomes these challenges for applications where a small amount of additional feedback is available. In the second part, I’ll discuss an offline evaluation method that requires no additional feedback and is therefore much more broadly applicable. Both methods come with strong theoretical and empirical properties.