Subspace-Based Decision Tree Retraining

In this study, we propose a method for generating additional learning samples using subspaces to enhance the discriminative performance of Decision Trees. Although Decision Trees are known for their high interpretability as classifiers, they are often limited in their discriminative performance. One approach to improve classifier performance is to increase the number of training samples. Inspired by this concept, previous studies have attempted to improve the performance of Decision Trees by generating samples with methods such as kernel-based approaches and labeling these samples with high-performance classifiers like Multi-Layer-Perceptrons. However, this method does not perform well for high-dimensional data. In expanded spaces, the probability of generating unsuitable samples for training within the entire space increases. To address this issue, we propose a method to generate additional learning samples within subspaces spanned by an extended k-subspace method, which divides the space into k subcategories. These additional samples are then incorporated into the decision tree's training process. We evaluated the effectiveness of our sample generation method using handwritten digit image data. As a result, we achieved an approximately 3% improvement in decision tree accuracy by adding 12 million generated training samples.