Abstract
This article proposes an Adversarial Prototype Decomposition (APD) algorithm as a method for improving the scalability of the Support Vector Machine (SVM) classifier. The algorithm partitions large datasets into smaller, more manageable subsets called regions, enabling the application of SVM classification to datasets containing hundreds of thousands of samples. A series of experiments were conducted on medium- and large-scale datasets to evaluate the approach. For medium-sized datasets (up to 60,000 samples), the algorithm maintains the same predictive performance as the reference SVM while accelerating the training and prediction phases by an average factor of 2.5. On large-scale datasets containing 500,000 samples, the proposed APD-SVM solution achieves significantly higher performance than the reference Cascade SVM, while simultaneously speeding up the training process by 1.7 times and the prediction phase by up to 36 times. Furthermore, the experiments demonstrate the importance of the Learning Vector Quantization (LVQ) procedure in optimizing the prototypes within the APD framework.
Paper Type
Full Paper
DOI
10.62036/ISD.2026.95
Scaling Machine Learning via Adversarial Prototype Decomposition: An SVM Case Study
This article proposes an Adversarial Prototype Decomposition (APD) algorithm as a method for improving the scalability of the Support Vector Machine (SVM) classifier. The algorithm partitions large datasets into smaller, more manageable subsets called regions, enabling the application of SVM classification to datasets containing hundreds of thousands of samples. A series of experiments were conducted on medium- and large-scale datasets to evaluate the approach. For medium-sized datasets (up to 60,000 samples), the algorithm maintains the same predictive performance as the reference SVM while accelerating the training and prediction phases by an average factor of 2.5. On large-scale datasets containing 500,000 samples, the proposed APD-SVM solution achieves significantly higher performance than the reference Cascade SVM, while simultaneously speeding up the training process by 1.7 times and the prediction phase by up to 36 times. Furthermore, the experiments demonstrate the importance of the Learning Vector Quantization (LVQ) procedure in optimizing the prototypes within the APD framework.
Recommended Citation
Dąbrowski, D., Blachnik, M. & Kopeć, G.(2026). Scaling Machine Learning via Adversarial Prototype Decomposition: An SVM Case Study. In M. Valenta, B. Mannová, R. Pergl, A. Przybylek, M. Lang, H. Linger, C. Schneider, N. Iivari, & E. Insfran (Eds.), Making ISD Sustainable: Reloaded with AI and Automation (ISD2026 Proceedings). Prague, Czech Republic: Czech Technical University in Prague. ISBN: 978-80-01-07585-2. https://doi.org/10.62036/ISD.2026.95