作者: Richard Nock , Frank Nielsen
DOI:
关键词:
摘要: We present a novel class of divergences induced by smooth convex function called total Jensen divergences. Those are invariant construction to rotations, feature yielding regularization ordinary conformal factor. analyze the relationships between this and recently introduced Bregman then proceed defining centroids as average distortion minimizers, study their robustness performance outliers. Finally, we prove that k-means++ initialization bypasses explicit centroid computations is good enough in practice guarantee probabilistically constant approximation factor optimal k-means clustering.