A Better Decision Tree: The Max-Cut Decision Tree with Modified PCA Improves Accuracy and Running Time

Abstract Decision trees are a widely used method for classification, both alone and as the building blocks of multiple different ensemble learning methods. The Max Cut decision tree introduced here involves novel modifications to standard, baseline variant classification tree, CART Gini. One modification an alternative splitting metric, Cut, based on maximizing distance between all pairs observations that belong separate classes sides threshold value. other modification, Node Means PCA, selects feature from linear combination input features constructed using adjustment principal component analysis (PCA) locally at each node. Our experiments show this node-based, localized PCA with metric can dramatically improve accuracy while also significantly decreasing computational time compared Gini tree. These improvements most significant higher-dimensional datasets. For example dataset CIFAR-100, enabled 49% improvement in accuracy, relative Gini, providing $$6.8 \times$$ 6.8 × speed up Scikit-Learn implementation expected advance capabilities difficult tasks.