Heavy-Tailed Neighbor Retrieval Visualizer
IEEE Access · 2026 · 14 · 67518-67534
DOI: 10.1109/access.2026.3687810 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Mouad Elhamdi und 3 weiteren Autor:innen. Die Arbeit nennt 36 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Smart Data Analysis Systems (SDAS) Research Group.
Abstract (Originalsprache)
Neighbor embedding methods like $t$ -SNE and UMAP are popular for visualizing high-dimensional data, but lack independent control over local versus global structure. The Neighbor Retrieval Visualizer (NeRV) addresses this by explicitly parameterizing the precision-recall trade-off. However, its reliance on Gaussian kernels makes it susceptible to the crowding problem, causing points to cluster centrally and reducing separation in dense areas. In this work, we introduce the here-termed heavy-tailed NeRV (HTNeRV), a scalable framework that resolves this limitation by integrating a Student- $t$ kernel with adjustable degrees of freedom into the NeRV objective. By introducing tail heaviness as a separate parameter, our method allows for simultaneous control over both cluster separation and the precision-recall balance. We validate this theoretical extension with a scalable, GPU-accelerated implementation and comprehensive benchmarks on image and single-cell RNA sequencing datasets. Our results, supported by rank-based criteria, show that HTNeRV remains competitive with strong baselines while offering a controllable trade-off between local neighborhood fidelity, class separability, and multiscale rank preservation.
Förderung
- Smart Data Analysis Systems (SDAS) Research Group
Zitieren
APA
Elhamdi, M., Aalaila, Y., Amhoud, E. M., & Peluffo-Ordóñez, D. H. (2026). Heavy-Tailed Neighbor Retrieval Visualizer. IEEE Access, 14, 67518-67534. https://doi.org/10.1109/access.2026.3687810
Vancouver
Elhamdi M, Aalaila Y, Amhoud EM, Peluffo-Ordóñez DH. Heavy-Tailed Neighbor Retrieval Visualizer. IEEE Access. 2026;14:67518-67534. doi:10.1109/access.2026.3687810
BibTeX
@article{Elhamdi2026,
author = {Elhamdi, Mouad and Aalaila, Yahya and Amhoud, El Mehdi and Peluffo-Ordóñez, Diego Hernán},
title = {Heavy-Tailed Neighbor Retrieval Visualizer},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {67518--67534},
doi = {10.1109/access.2026.3687810}
}
EUVIMED