Section: New Results

Subspace clustering

Participants: S Peignier, C Rigotti

We developed an algorithm to tackle the subspace clustering problem over a data stream containing clusters than change over time. Very few subspace clustering algorithms can handle such streams. Our starting point was the work made in the team on evolution of evolution mechanisms and on a preliminary bio-inspired algorithm that we have proposed last year. This previous algorithm included many bio-like features like variable genome length and organization, functional and non-functional elements, and variation operators including chromosomal rearrangements. It achieved satisfying results on standard benchmark data sets but was not designed to process dynamic streams. The new algorithm finds and adapts changing clusters over such streams, while preserving high cluster quality. It has been successfully used to build the evolving music generation system EvoMove.