Home | Publications | BKS19

Rock - Let the Points Roam to Their Clusters Themselves

MCML Authors

Anna Beer

Dr.

* Former Member

→ Group Thomas Seidl
Database Systems, Data Mining and AI

Daniyal Kazempour

Dr.

* Former Member

→ Group Thomas Seidl
Database Systems, Data Mining and AI

Thomas Seidl

Prof. Dr.

Director

Database Systems, Data Mining and AI

Abstract

In this work we present Rock, a method where the points roam to their clusters using k-NN. Rock is a draft for an algorithm which is capable of detecting non-convex clusters of arbitrary dimension while delivering representatives for each cluster similar to, e.g., Mean Shift or k-Means. Applying Rock, points roam to the mean of their k-NN while k increments in every step. Like that, rather outlying points and noise move to their nearest cluster while the clusters themselves contract first to their skeletons and further to a representative point each. Our empirical results on synthetic and real data demonstrate that Rock is able to detect clusters on datasets where either mode seeking or density-based approaches do not succeed.

inproceedings