<p>Instance selection is a data preprocessing method in data mining that aims to reduce the volume of the training dataset. Reducing samples from a large dataset offers benefits such as lower storage requirements, reduced computational costs, increased processing speed, and, in some cases, improved accuracy for learning algorithms. However, reducing samples from large datasets is also a challenging task due to their sheer volume. Recently, numerous instance selection methods for big data have been proposed, often facing challenges such as low accuracy and slow processing speed. In this research, we propose a fast and efficient three-step method based on the divide-and-conquer approach. In the first step, the training set is divided based on the number of classes. Next, representative summaries of each class are extracted. Finally, samples from each class are reduced independently while considering the representatives of other classes. By using a proposed ranking-based method, it is possible to accurately identify less important and noisy samples. For a comprehensive evaluation, we utilized 20 well-known large datasets and three synthetic datasets featuring challenging structures. The results demonstrate the superiority of the proposed method over four recent related methods.</p>