Feature-weighted earth mover’s distance for enhanced k -means clustering: a multi-dataset evaluation on clinical data
Machine Learning: Science and Technology · 01.06.2026 · 7 (3) · 035028
DOI: 10.1088/2632-2153/ae6bfe ↗Kurz gefasst
Fachartikel in der Zeitschrift Machine Learning: Science and Technology, online erschienen am 22.05.2026. Verfasst von Shanshan Fang und 1 weiteren Autor:innen. Die Arbeit nennt 58 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Clustering methods based on K ‐means typically rely on Euclidean distance, which limits their ability to capture distributional structure in heterogeneous data. To address this limitation, we integrate optimal transport (OT) theory into K ‐means clustering through the earth mover’s distance (EMD) and introduce EMD + , a feature‐weighted extension that incorporates data‐driven feature attribution while preserving the probabilistic properties of the input. Building on this foundation, we extend it to a family of feature-weighted distance functions that leverage six data-driven attribution strategies, including permutation feature importance and principal component analysis. We conduct extensive experiments across five real-world datasets to evaluate clustering performance using a comprehensive suite of internal and external validation metrics. Results show that EMD consistently improves clustering quality over Euclidean distance across all algorithms. Moreover, EMD + offers additional gains in specific scenarios, particularly for classical and Mini‐Batch K ‐means. Aggregated results across all datasets and metrics demonstrate a robust advantage of OT‐based distances relative to Euclidean baselines. These findings demonstrate that integrating OT and feature attribution into clustering algorithms can significantly enhance their robustness and effectiveness when applied to real-world, heterogeneous datasets. This work provides a scalable, interpretable, and domain-adaptive framework for advancing clustering methodologies in high-stakes applications such as health informatics and decision support systems. Although our evaluation uses medical datasets, the approach is general and applies to other domains with standard preprocessing.
Zitieren
APA
Fang, S., & Uddin, S. (2026). Feature-weighted earth mover’s distance for enhanced k -means clustering: a multi-dataset evaluation on clinical data. Machine Learning: Science and Technology, 7(3), 035028. https://doi.org/10.1088/2632-2153/ae6bfe
Vancouver
Fang S, Uddin S. Feature-weighted earth mover’s distance for enhanced k -means clustering: a multi-dataset evaluation on clinical data. Machine Learning: Science and Technology. 2026;7(3):035028. doi:10.1088/2632-2153/ae6bfe
BibTeX
@article{Fang2026,
author = {Fang, Shanshan and Uddin, Shahadat},
title = {Feature-weighted earth mover’s distance for enhanced k -means clustering: a multi-dataset evaluation on clinical data},
journal = {Machine Learning: Science and Technology},
year = {2026},
volume = {7},
number = {3},
pages = {035028},
doi = {10.1088/2632-2153/ae6bfe}
}
EUVIMED