Determining the k in k-means with MapReduce

Thibault Debatty (1, 2) , Pietro Michiardi (2) , Wim Mees (1) , Olivier Thonnard (3)

1 Ecole Royale Militaire
2 Eurecom [Sophia Antipolis]
3 Symantec Research Labs

In this paper we propose a MapReduce implementation of G-means, a variant of k-means that is able to automatically determine k, the number of clusters. We show that our implementation scales to very large datasets and very large values of k, as the computation cost is proportional to nk. Other techniques that run a clustering algorithm with different values of k and choose the value of k that provides the " best " results have a computation cost that is proportional to nk 2. We run experiments that confirm that the processing time is proportional to k. These experiments also show that, because G-means adds new centers progressively, if and where they are needed, it reduces the probability to fall into a local minimum, and finally finds better centers than classical k-means processing.

Domaines: Algorithme et structure de données [cs.DS]; Intelligence artificielle [cs.AI]; Calcul parallèle, distribué et partagé [cs.DC]

Thibault Debatty, Pietro Michiardi, Wim Mees, Olivier Thonnard. Determining the k in k-means with MapReduce. EDBT/ICDT 2014 Joint Conference, Mar 2014, Athènes, Greece. ⟨hal-01525708⟩