One-class support vector machine (OCSVM) is a widely used tool for credit risk detection, yet its traditional hinge loss exhibits inadequate robustness against outliers and noise, resulting in suboptimal performance in practical credit risk evaluation. To mitigate these drawbacks, this paper integrates the 0-1 loss into the OCSVM framework (termed \(L_{0/1}\) -OCSVM), building on Sch \(\ddot{o}\) lkopf’s classical OCSVM and the representation theorem. Experiments on 8 raw and artificial credit risk data sets show that the optimized model outperforms three baseline OCSVM variants in F1 score and G-mean for most data sets, while requiring the minimum number of support vectors across all data sets. Further tests with injected feature noise and outliers, as well as parameter insensitivity analysis, confirm its superior robustness and stable generalization capability. These findings validate that the proposed model achieves balanced performance in identifying normal and high-risk users, with practical deployment advantages.
