Adaptive-threshold max-SVM model with progressive-iterative algorithm for high-dimensional data classification and feature selection

Abstract

Purpose Support vector machines have been progressively refined in existing research, but there are still certain limitations in feature selection for high-dimensional data. Design/methodology/approach This paper proposes an adaptive-threshold max-SVM model, which seeks to improve the accuracy and effectiveness of classification and identifies key features. The model comprises a set of SVM sub-models, each trained on a specific feature combination and adjusting its decision boundary with adaptively optimized thresholds. All sub-models independently generate classification results, and the final decision is determined by selecting the maximum value from these outputs. Findings In biomedical data analysis, this model achieves high accuracy with 100\% sensitivity. It delivers superior performance over conventional models in repeated cross-validation and exhibits strong generalization and robustness, which are also validated on non-biomedical datasets. Originality/value This paper designs a novel adaptive-threshold max-SVM model equipped with the Progressive-Iterative Algorithm. The algorithm consists of a three-step screening and iterative construction process that identifies an optimal combination of sub-models from a large candidate pool. It enables efficient feature selection, delivering favorable classification performance, generalization ability and robustness.