Weight-based search to find clusters around medians in subspaces

Sergio Peignier 1, 2 Christophe Rigotti 3, 1, 2 Anthony Rossi 1 Guillaume Beslon 1, 2
1 BEAGLE - Artificial Evolution and Computational Biology
LIRIS - Laboratoire d'InfoRmatique en Image et Systèmes d'information, Inria Grenoble - Rhône-Alpes, LBBE - Laboratoire de Biométrie et Biologie Evolutive - UMR 5558
3 DM2L - Data Mining and Machine Learning
LIRIS - Laboratoire d'InfoRmatique en Image et Systèmes d'information
Abstract : There exist several clustering paradigms, leading to different techniques that are complementary in the analyst toolbox, each having its own merits and interests. Among these techniques, the K-medians approach is recognized as being robust to noise and outliers, and is an important optimization task with many different applications (e.g., facility location). In the context of subspace clustering, several paradigms have been investigated (e.g., centroid-based, cell-based), while the median-based approach has received less attention. Moreover, using standard subspace clustering outputs (e.g., centroids, medoids) there is no straightforward procedure to compute the cluster membership that optimizes the dispersion around medians. This paper advocates for the use of median-based subspace clustering as a complementary tool. Indeed, it shows that such an approach exhibits satisfactory quality clusters when compared to well-established paradigms, while medians have still their own interests depending on the user application (robustness to noise/outliers and location optimality). This paper shows that a weight-based hill climbing algorithm using a stochastic local exploration step can be sufficient to produce the clusters.
Document type :
Conference papers
Complete list of metadatas

Cited literature [18 references]  Display  Hide  Download

https://hal.archives-ouvertes.fr/hal-01869974
Contributor : Christophe Rigotti <>
Submitted on : Friday, September 7, 2018 - 9:30:24 AM
Last modification on : Wednesday, May 15, 2019 - 2:22:02 PM
Long-term archiving on : Saturday, December 8, 2018 - 12:50:50 PM

File

prelim_subcmedians.pdf
Files produced by the author(s)

Identifiers

  • HAL Id : hal-01869974, version 1

Citation

Sergio Peignier, Christophe Rigotti, Anthony Rossi, Guillaume Beslon. Weight-based search to find clusters around medians in subspaces. SAC 2018 - ACM Symposium On Applied Computing, Apr 2018, Pau, France. pp.1-10. ⟨hal-01869974⟩

Share

Metrics

Record views

189

Files downloads

119