Work

A Relative Probabilistic Density Weighted Granular-Ball Twin Support Vector Machine for Robust Classification

Year: 2026
Type: preprint
Abstract: Classical Twin Support Vector Machine (TWSVM) uses individual samples as input and constructs two nonparallel hyperplanes for binary classification. To improve computational efficiency, Granular-Ball Twin Support Vector Machine (GBTWSVM) replaces point-level input with granular-ball-level input. However, most existing granular-ball generation methods based on virtual centers may fail to accurately represent the original data distribution on large-scale and complex datasets. In addition, existing relative-density weighting methods mainly operate at the point level, whereas distribution-aware weighting at the granular-ball-center level has not been sufficiently investigated. To address these problems, this paper proposes a Twin Support Vector Machine based on Local Density Granular-Ball Generation and Relative Probabilistic Density weighting, termed RPDLDGBTWSVM. The proposed method selects real samples with relatively high local densities as granular-ball centers and determines granular-ball radii according to within-class compactness and between-class separability. A KLIEP-based relative probabilistic density estimation method is introduced at the granular-ball-center level. The resulting weights are incorporated into the slack-variable penalty terms, while the granular-ball radii are introduced into the classification constraints. Theoretical analysis shows that the approximation error between the granular-ball-center-level RPD-weighted loss and the corresponding point-level weighted loss is controlled by the granular-ball radii. The resulting subproblems remain convex and feasible. The proposed model can also be extended to nonlinear classification through kernel methods. Comparative experiments are conducted on multiple UCI datasets and medical binary classification datasets. The experimental results show that the proposed method achieves competitive overall performance in terms of classification accuracy and computational efficiency.
Institution: Jinan University
Language: English
Open Access status: green
OpenAlex ID: W7215121973
Cites: 35
Cited by:
Related to: 9