An Innovative Subsampling Approach for Efficient SVM Training with Large Datasets

Support vector machines (SVMs) are widely recognized for their effectiveness in handling classification problems, owing to their solid theoretical foundation and excellent generalization performance. However, despite these advantages, SVMs have a significant drawback in the form of high computational time, which increases with the size of the training dataset. To address this limitation, this article presents a novel adaptive sequential subsampling method designed to accelerate the training process of SVMs. The proposed method consists of two stages. In the first stage, a space-filling design is employed to group samples into cells. Then, an initial pilot SVM model is trained by utilizing the centroids and corresponding labels of these cells. In the second stage, an adaptive sequential stratified sampling method, based on the distance between each cell and the hyperplane, is employed to select informative samples, thereby enhancing the SVM model. Numerical studies show that our approach achieves classification accuracy that is comparable to or even better than that of basic SVM, while requiring only approximately 1% of the CPU time. Consequently, our algorithm is a more efficient choice for large-scale data applications.