Incremental Fuzzy Decision Tree Based on Cumulative Mutual Information

Abstract Traditional decision trees and fuzzy decision trees (FDTs) face inherent challenges in managing uncertainty, noise, and evolving data streams. While FDTs improve upon classical decision trees by incorporating fuzzy logic to handle ambiguity, they remain static and cannot adapt to changes in data distribution over time. This paper introduces the Dynamic Fuzzy Decision Tree (DFDT), an algorithm that combines fuzzy logic with incremental learning to address these limitations. DFDT leverages cumulative mutual information to guide its dynamic restructuring, enabling it to update its model continuously while preserving its ability to reason under uncertainty. It integrates attribute swapping and branch pruning mechanisms to maintain a compact, interpretable structure without incurring significant computational overhead. Experimental evaluations on benchmark datasets show that DFDT consistently outperforms static FDTs in dynamic environments, achieving, for example, a 3.6% improvement in accuracy on the LED-Drift dataset and ranking highest in 10 out of 11 evaluated cases. The single-pass DFDT achieves competitive performance with minimal resource usage, making it suitable for real-time data stream scenarios. In contrast, its epoch-based counterpart offers a balance between incremental learning and batch-level accuracy. Compared to traditional incremental learning algorithms, DFDT achieves comparable or better performance while inherently handling uncertainty via fuzzy logic. This work advances the development of adaptive classifiers that maintain interpretability and computational efficiency, positioning DFDT as a strong candidate for real-time applications in dynamic and uncertain environments.