Python Data Analysis Cookbook

Comparing results with a dummy classifier

The scikit-learn DummyClassifier class implements several strategies for random guessing, which can serve as a baseline for classifiers. The strategies are as follows:

• stratified: This uses the training set class distribution
• most_frequent: This predicts the most frequent class
• prior: This is available in scikit-learn 0.17 and predicts by maximizing the class prior
• uniform: This uses an uniform distribution to randomly sample classes
• constant: This predicts a user-specified class

As you can see, some strategies of the DummyClassifier class always predict the same class. This can lead to warnings from some scikit-learn metrics functions. We will perform the same analysis as we did in the Computing precision, ...