ProDiSE: A proximity and distance scoring based pruning technique for large-scale supervised learning

Abstract

In the current era of data-driven computing, efficiently handling large-scale datasets has become a fundamental necessity. However, most existing supervised learning models struggle to scale effectively due to their high computational demands. While existing models achieve good performance, their scalability is often limited, motivating the need for efficient data reduction strategies. In this work, we propose a non-iterative pruning framework, termed proximity and distance scoring ensemble (ProDiSE), with the goal of selecting a compact yet representative subset of training samples for scalable supervised learning. ProDiSE evaluates sample representativeness using two complementary geometric measures computed over random subsamples: a local distance score (LDS), which captures neighborhood density, and a global compactness score (GCS), which reflects proximity to class centroids. These scores are adaptively combined using a variance-based weighting scheme, allowing the informative samples to dominate for each class. To improve robustness against randomness, the scoring process is repeated across multiple ensemble runs, and final representativeness is determined via a 75th percentile aggregation strategy. The proposed ProDiSE framework is integrated with the least squares support vector machine (LSSVM), yielding ProDiSE-LSSVM, a sparse and computationally efficient classifier. The resulting model exhibits linear time and memory complexity with respect to the training set size. Extensive experiments on a wide range of real-world datasets demonstrate that ProDiSE-LSSVM consistently achieves higher classification accuracy than competing models while substantially reducing training cost. Overall, the proposed approach provides an effective and scalable solution for large-scale classification, offering a favorable balance between performance and computational efficiency. � 2026 Elsevier B.V.