A Novel Nonparallel Hyperplane Based Extreme Learning Machine for Classification

Recently, researchers have developed nonparallel variants of both extreme learning machine (ELM) and support vector machine (SVM) classifiers to improve learning speed and generalization. However, these nonparallel methods typically solve two SVM-type optimization problems. Thus far, no existing nonparallel ELM variants introduce nonparallel hyperplanes into the traditional ELM framework or use a unified optimization problem. To address this, the paper proposes a nonparallel hyperplane based extreme learning machine (NHELM), which solves a unified ELM-type optimization problem. The proposed nonparallel ELM employs nonparallel hyperplanes to improve the traditional ELM for better generalization. In NHELM, nonparallel hyperplanes are designed to align with class-specific data distributions to construct a more effective angle bisector (separating hyperplane) that improves class separation. Experiments on 20 KEEL binary benchmark datasets and six MedMNIST medical image datasets using both linear and gaussian kernels show that NHELM achieves up to a 3.4% improvement in classification performance over traditional ELM, and a 2.2% improvement over existing nonparallel ELM variants, based on the averages of metrics such as accuracy, G-mean, and area under the curve (AUC). The results demonstrate that the proposed method improves pattern learning and generalization, making it highly useful for diverse real-world classification problems.