Analytical parameter selection with critical computational resources

Ramírez M.; Luyo J.; Luna A.; Machuca L.; Chong M. · Issues in Information Systems (2026) · DOI: 10.48009/4_iis_2026_117

Resumen

Support Vector Machines (SVMs) are essential for machine learning classification, yet their performance depends heavily on Gaussian Kernel hyperparameter selection. Traditional iterative tuning methods, such as cross-validation and heuristics, suffer from high computational complexity, limiting real-time application under constrained resources. This study addresses these limitations by proposing a geometric-differentiable framework for analytical parameter selection based on class separability. By incorporating direct and inverse variation criteria governed by a non-linear control exponent, the proposed approach allows instantaneous, closed-form adjustments of optimization weights. Algorithmic scalability was validated using the large-scale ijcnn1 benchmark, achieving a 99.8% reduction in hyperparameter selection time compared to conventional Grid Search. Additionally, the Statlog (German Credit Data) dataset was utilized as an analytical proxy to simulate socioeconomic capacity assessments within post-crisis environments. Results show that the proposed analytical criteria maintain high classification accuracy, reaching up to 98.8% in testing sets. This mathematical optimization resolves the conflict between computational speed and operational flexibility, establishing a plausible, highly efficient framework to enhance decision-support systems when processing power is critically limited.

Instituciones: Universidad Nacional Mayor de San Marcos; Universidad del Pacífico