Generalized bell fuzzy proximal twin support vector machine for multi-class classification

Nikita Grewal, Yash Arora, Shiv Kumar Gupta, Sanjeev Kumar

Abstract: Fuzzy support vector machine is a versatile machine learning technique for addressing binary classification and regression challenges. However, it still faces problems related to borderline noise, outliers, and multi-class issues. To handle these problems, we introduce a novel technique called generalized bell fuzzy proximal twin support vector machine for multi-category classification tasks. The generalized bell function gives lower membership values to training samples farther from the class center, reducing the effect of borderline noise and outliers in the construction of the hyperplane. We employ this membership function in the proximal twin support vector machine, which enhances the computational efficiency of the model, as it requires solving only a system of linear equations that involves matrix inversion. We used a one-against-all approach to solve the multi-class problem. Further, the performance of the proposed method is assessed in comparison to baseline models by performing the experiments on ten benchmark datasets. The evaluation is conducted using both linear and Gaussian kernels, considering metrics such as F-measure, accuracy, computational time, and G-mean. We also performed comprehensive statistical tests to validate our findings. The results highlight the effectiveness of the proposed technique in comparison to existing methods, with an impressive mean accuracy of 92.36% with linear and 93.18% with the Gaussian kernel, demonstrating its potential to address classification challenges in real-world problems effectively.