SOTAI: Guaranteed ML Behaviors

When training ML models, it is never the case that your data is perfect or infinite. The noise in your data will result in an ML model that has pockets of unexplainable behavior that go directly against your expectations. For example, consider training a model to predict the price of a house. You generate a prediction only to realize you input the wrong square footage, but after increasing the square footage the predicted price goes down. Now you struggle to trust the results of your model and are back to the drawing board. SOTAI enables you to train models that guarantee certain real-world expectations such as this one. Simply set a constraint such that increasing square footage should only ever increase the predicted price, all else being equal, and your model will guarantee this behavior even on unseen examples. Of course, it is only the expected increase that is guaranteed – the amount of increase is learned from your data.