Online evaluation of machine learning models

Description

Academic machine learning involves almost exclusively offline evaluation of machine learning models. In the real world this is, somewhat surprisingly, often only good enough for a rough cut that eliminates the real dogs. For production work, online evaluation is often the only option to determine which of several final-round candidates might be chosen for further use. As Einstein is rumored to have said, theory and practice are the same, in theory. In practice, they’re different. So it is with models. Part of the problem is interaction with other models and systems. And part of the problem has to do with variability of the real world. Often there are adversaries at work. It may even be sunspots. One particular problem arises when models choose their own training data and thus couple back onto themselves.

In addition to these difficulties, production models almost always have service-level agreements that have to do with how quickly they must produce results and how often they are allowed to fail. These operational considerations can be as important as the accuracy of the model…right results returned late are worse than slightly wrong results returned in time.

Ted Dunning explores a survey of useful ways to evaluate models in real-world use, including decoy and canary models, nonlinear latency histogramming, model-delta diagrams, and more. These techniques may sound arcane, but each has a simple heart and should not require any advanced mathematics to understand.