At each node, the training data branches (splits) into two subsets, divided by the feature with most-information about the target variable. High interpretability at the risk of overfitting.