Random Machines (RM) are ensemble models that combine bootstrap sampling with support vector machines (SVMs) trained using randomly sampled kernels. Prior work has shown that RM can outperform Random Forests on many classification and regression tasks, but also that performance deteriorates in two practically important regimes: strong class imbalance and high-dimensional, low-sample-size data. In this paper we propose Cost-Sensitive Sparse Random Machines (CS–SRM), an extension of RM designed specifically for these settings. Each base learner in CS–SRM is a cost-sensitive SVM trained on a random feature subspace with either sparse linear or nonlinear kernels, and ensemble weights are derived from imbalance-aware out-of-bag metrics such as Matthews correlation coefficient and \(F_1\)-score. We outline the methodology, describe a simulation and case-study evaluation design, and summarize the types of results such a study would yield. Conceptually, CS–SRM improves minority-class performance and stability in high-dimensional regimes while preserving the flexibility of multi-kernel ensembles.
Random Machines (RM) are ensemble models that combine bootstrap sampling with support vector machines (SVMs) trained using randomly sampled kernels. Prior work has shown that RM can outperform Random Forests on many classification and regression tasks, but also that performance deteriorates in two practically important regimes: strong class imbalance and high-dimensional, low-sample-size data. In this paper we propose Cost-Sensitive Sparse Random Machines (CS–SRM), an extension of RM designed specifically for these settings. Each base learner in CS–SRM is a cost-sensitive SVM trained on a random feature subspace with either sparse linear or nonlinear kernels, and ensemble weights are derived from imbalance-aware out-of-bag metrics such as Matthews correlation coefficient and \(F_1\)-score. We outline the methodology, describe a simulation and case-study evaluation design, and summarize the types of results such a study would yield. Conceptually, CS–SRM improves minority-class performance and stability in high-dimensional regimes while preserving the flexibility of multi-kernel ensembles.
