Humboldt-Universität zu Berlin - High Dimensional Nonstationary Time Series


K-expectiles clustering

Bingling Wang
Yingxing Li
Wolfgang Karl Härdle

K-means clustering is one of the most widely-used partitioning algorithm in
cluster analysis due to its simplicity and computational efficiency, but it may
not provide ideal clustering results when applying to data with non-spherically
shaped clusters. By considering the asymmetrically weighted distance, We propose
the K-expectile clustering and search the clusters via a greedy algorithm that
minimizes the within cluster τ-variance. We provide algorithms based on two
schemes: the fixed τ clustering, and the adaptive τ clustering. Validated by
simulation results, our method has enhanced performance on data with asymmetric
shaped clusters or clusters with a complicated structure. Applications of our
method show that the fixed τ clustering can bring some flexibility on
segmentation with a decent accuracy, while the adaptive τ clustering may yield
better performance. All calculation can be redone via

clustering, expectiles, asymmetric quadratic loss, image segmentation

JEL Classification: