|
Abstract:
|
When using a graphical tool, such as a prediction profiler, to explore predictions of a model, it can often be difficult to tell if predictions are extrapolating too far outside of the original data. Additionally, without proper constraints, optimizing over the prediction surface of a model can often lead to extrapolated solutions that are of no practical use. This talk explores methods that can be used in a graphical tool to let users know when prediction points are extrapolated too far outside of the original data and can also be used to perform constrained optimization to prevent extrapolated optimal solutions. Ideally, a method would work well on any type of statistical or machine learning model. We will demonstrate this method on least squares models and discuss research to extend these methods to general machine learning models.
|
