Section: New Results

Subspace clustering based on medians using evolutionary algorithms

Participants: Sergio Peignier, Christophe Rigotti, Jonas Abernot, Guillaume Beslon

Subspace clustering is a data mining task that searches for objects sharing similar features, and at the same time looks for the subspaces where these similarities appear. For this reason subspace clustering is recognized as more general and complicated than standard clustering, since it needs to detect these relevant subspaces. Taking advantage of the expertise of the team in evolution in silico, we previously showed that evolutionary algorithms are promising approaches to address this problem. Another important clustering task is the K-medians one, where objects are grouped around medians, leading to cluster centers more robust to noise and outliers. In order to take advantage of these benefits within the subspace clustering process it-self, we developed a new evolutionary algorithm, KymeroClust, that builds cluster centers that are medians in subspaces. This algorithm takes advantage of an evolvable representation of the genotypes to adapt the numbers of clusters produced and the subspace dimensionalities. It is based on new bio-inspired mutation operators to evolve the cluster centers as medians and is able to handle streaming data. KymeroClust has been compared to the main subspace clustering methods and turns out to be very competitive both in terms of cluster quality and runtime, while requiring an easier parameter setting.

Publications: [12], [24]