A Unified Instance Selection Approach Based on Distributional Similarity for Machine Learning

Hüseyin Tunç

In this study, we propose a unified instance selection approach, applicable to both regression and classification tasks, that selects a reduced dataset whose distribution matches that of the original, thereby preserving its key statistical and structural properties. Our central contribution is to analytically show that the instance selection problem, aiming to ensure distributional similarity measured via the Wasserstein distance, is equivalent to the balanced k-medoid problem when the compression ratio is integer-valued. We further extend this equivalence to arbitrary (non-integer) ratios with explicit bounds on the resulting approximation error. This equivalence allows the instance selection problem to be addressed directly with existing balanced k-medoid solvers. To this end, we employ a Lloyd-style alternating (EM-type) constrained k-medoid algorithm, obtained by extending an existing constrained clustering method. Extensive experiments across binary classification, multi-class classification, and regression tasks confirm the method's effectiveness: it significantly outperforms the task-specific baseline at most reduction ratios on all seven multi-class datasets, outperforms the boundary-focused baseline on the majority of binary settings—most clearly on the higher-dimensional datasets—and attains lower error on four of five regression datasets.

Short Bio

Hüseyin Tunç has served as an Associate Professor of Operations Research in the Department of Management at the Social Sciences University of Ankara (SSUA) since 2019. He previously worked as an Assistant Professor at Hacettepe University. Hüseyin has also held visiting and postdoctoral appointments at the Amsterdam Business School, EPFL in Lausanne, and Mississippi State University. His research focuses on the theory and application of operations research for optimization under uncertainty, with a focus on data analytics, data-driven decision-making, supply chain management, and inventory control.

Venue

Friday, November 13, 2026, 4.00 pm

IE Building, Halim Doğrusöz Auditorium (IE 03)