Saved in:
Bibliographic Details
Main Authors: de Mathelin, Antoine, Cecchi, Nicolas Enrique, Deheeger, François, Mougeot, Mathilde, Vayatis, Nicolas
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2501.19285
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866916592518955008
author de Mathelin, Antoine
Cecchi, Nicolas Enrique
Deheeger, François
Mougeot, Mathilde
Vayatis, Nicolas
author_facet de Mathelin, Antoine
Cecchi, Nicolas Enrique
Deheeger, François
Mougeot, Mathilde
Vayatis, Nicolas
contents This paper proposes a novel k-medoids approximation algorithm to handle large-scale datasets with reasonable computational time and memory complexity. We develop a local-search algorithm that iteratively improves the medoid selection based on the estimation of the k-medoids objective. A single batch of size m << n provides the estimation, which reduces the required memory size and the number of pairwise dissimilarities computations to O(mn), instead of O(n^2) compared to most k-medoids baselines. We obtain theoretical results highlighting that a batch of size m = O(log(n)) is sufficient to guarantee, with strong probability, the same performance as the original local-search algorithm. Multiple experiments conducted on real datasets of various sizes and dimensions show that our algorithm provides similar performances as state-of-the-art methods such as FasterPAM and BanditPAM++ with a drastically reduced running time.
format Preprint
id arxiv_https___arxiv_org_abs_2501_19285
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle OneBatchPAM: A Fast and Frugal K-Medoids Algorithm
de Mathelin, Antoine
Cecchi, Nicolas Enrique
Deheeger, François
Mougeot, Mathilde
Vayatis, Nicolas
Machine Learning
This paper proposes a novel k-medoids approximation algorithm to handle large-scale datasets with reasonable computational time and memory complexity. We develop a local-search algorithm that iteratively improves the medoid selection based on the estimation of the k-medoids objective. A single batch of size m << n provides the estimation, which reduces the required memory size and the number of pairwise dissimilarities computations to O(mn), instead of O(n^2) compared to most k-medoids baselines. We obtain theoretical results highlighting that a batch of size m = O(log(n)) is sufficient to guarantee, with strong probability, the same performance as the original local-search algorithm. Multiple experiments conducted on real datasets of various sizes and dimensions show that our algorithm provides similar performances as state-of-the-art methods such as FasterPAM and BanditPAM++ with a drastically reduced running time.
title OneBatchPAM: A Fast and Frugal K-Medoids Algorithm
topic Machine Learning
url https://arxiv.org/abs/2501.19285