Abstract The challenge of imbalanced and small-sample data poses a fundamental constraint for classifiers in critical domains such as medical diagnosis, industrial fault detection and biometric recognition. As a result, models often fail to detect crucial minority samples, suffering from low Recall and reduced reliability. In this study, we propose an improved Unilateral Relaxed Support Vector Machine (UR-SVM) integrated with Convolutional Neural Networks (CNNs) for feature extraction. Asymmetric margin constraints are implemented, wherein a hard margin is applied to the minority class and a soft margin to the majority. This strategy directly improves minority-class recall without compromising the overall balance. Based on experiments across three datasets (UCI benchmarks, CIFAR-10, COVID-19 Radiography Dataset), our proposed CNN + UR-SVM offers significant advantages over traditional and cost-sensitive SVMs. The model achieves near-perfect Recall on small-sample datasets while maintaining competitive performance in Accuracy, F1-score, and G-mean, particularly when integrated with deep networks like ResNet and DenseNet. The result shows that the UR-SVM, as a robust solution to imbalance bias, is maintaining performance across varying data ratios and excelling under highly imbalanced scenarios. Overall, by achieving high Recall while preserving accuracy, the CNN + UR-SVM framework thus represents a reliable and transparent margin-based solution (with explicit unilateral constraints) for high-risk domains where missed detections must be minimized.
Jing et al. (Wed,) studied this question.