A split partitions the observations at a tree node using a feature and a dividing point. Classification trees choose splits that reduce class impurity; with entropy, this is expressed as information gain. Regression trees choose splits that reduce variation in the numerical outcome. section 9.1
Tree training is greedy: each split is selected for its immediate improvement, without looking ahead or revisiting earlier choices. The locally best split need not produce the best overall tree. section 9.1.1
Small changes in the training sample can change which candidate split is selected, especially when candidates perform similarly. A change near the root redirects observations into different branches, potentially changing the structure of the tree that follows. section 9.1.1



