The Supervised Learning Workflow

Scikit-learn ships a small medical dataset that most practitioners have loaded and few have examined: 442 diabetes patients, ten measurements taken at a baseline visit, and a number recording how far each patient's disease had progressed one year later. Fit a decision tree with no depth limit to 353 of those patients and its training error is exactly zero; it reproduces every one of them. Score it on the other 89 and its mean squared error is 6,997. A "model" that ignores all ten measurements and predicts the training patients' average progression for everyone scores 5,135 on the same 89 patients.