`

Timezone: »

 
Modal-sets, and density-based Clustering
Samory Kpotufe

Sat Dec 09 02:00 PM -- 02:30 PM (PST) @ None

Modes or Modal-sets are points or regions of space where the underlying data density is locally-maximal. They are relevant in problems such as clustering, outlier detection, or can simply serve to identify salient structures in high-dimensional data (e.g. point-cloud data from medical imaging, astronomy, etc).

In this talk we will argue that modal-sets, as general extremal surfaces, yield more stable clustering than usual modes (extremal points) of a density. For one, modal-sets can be consistently estimated, at non-trivial convergence rates, despite the added complexity of unknown surface-shape and dimension. Furthermore, modal-sets neatly dovetail into existing approaches that cluster data around point-modes, yielding competitive, yet more stable clustering on a range of real-world problems.

Author Information

Samory Kpotufe (Princeton University)

More from the Same Authors