A New Decision Tree Based on Intuitionistic Fuzzy Twin Support Vector Machines

Abstract

Effectively classifying anomalies in a multi-class setting holds significant importance in domains such as medical datasets, fraud detection, and anomaly detection. This task presents challenges that include efficient training on large datasets, accurate classification in imbalanced scenarios, and sensitivity to high imbalance ratios (IR). This paper introduces a novel approach, the Intuitionistic Fuzzy Twin Support Vector Machine-based Decision Tree (NDT-IFTSVM), aimed at addressing these issues. NDT-IFTSVM integrates IFTSVM and decision tree methodologies, offering an efficient solution for multi-class classification. The proposed algorithm constructs a decision tree comprised of a series of two-class IFTSVMs. To enhance balance and separability, the multi-class method iteratively divides into two sets based on distance between class centres and instance distribution. This recursive process continues until each subset exclusively contains a single class, facilitating effective classification. To handle highly imbalanced datasets, NDT-IFTSVM incorporates a rational weighting strategy. Additionally, we refine NDT-IFTSVM by introducing a regularization term that maximizes the margin between the bounding and proximal hyperplanes, mitigating the impact of noise and outliers. Finally, a coordinate descent system with shrinking by an active set is applied to reduce the computational complexity. Numerical evaluations employ the bootstrap technique with a 95% confidence interval and statistical tests to quantify the significance of performance improvements. Experimental results on 12 datasets demonstrate the efficacy of the proposed method, showcasing promising outcomes compared to other techniques documented in the literature.