For those who like or need to learn machine learning with scikit-learn, take a look at my tutorial series on this amazing package:

All images by creator.

Dummy models are very simplistic models that should be used as a baseline to check your actual models. A baseline is just a few type of reference point to check yourself to. If you compute your first cross-validation results to estimate your model’s performance, you often know that the upper the rating the higher, and if the rating is pretty high on the primary try, that’s great. But it surely isn’t often the case.

What to do if the primary accuracy rating is pretty low — or lower than what you’d want or expect? Is it due to the information? Is it due to your model? Each? How can we all know quickly if our model isn’t badly tuned?

Dummy models are here to reply these questions. Their complexity and “intelligence” are very low: the concept is that you could compare your models to them to see how a lot better you might be than the “stupidest” models. Note that they don’t intentionally predict silly values, they simply take the simplest, very simplistic smart guess. For those who model gives worst performance than the dummy model, it’s best to tune or change your model completely.

A straightforward example for a dummy regressor can be to at all times predict the mean value of the training goal, regardless of the input: it’s not ideal, but on average it gives an inexpensive simplistic guess. In case your actual model gives worse results than this very, quite simple approach, it is advisable to review your model.