Fast and Provably Good Seedings for k-Means Olivier Bachem, Mario Lucic, S. Hamed Hassani, Andreas Krause F a s t a n d P r o v a b l y G o o d S e e d i n g s f o r k - M e a n s Teaser F a s t a n d P r o v a b l y G o o d S e e d i n g s f o r k - M e a n s Teaser 1'064x 1.32% @ UP TO SPEEDUP RELATIVE ERROR COMPARED TO K-MEANS++ F a s t a n d P r o v a b l y G o o d S e e d i n g s f o r k - M e a n s Teaser 1'064x 1.32% @ UP TO SPEEDUP RELATIVE ERROR COMPARED TO K-MEANS++ + THEORETICAL GUARANTEES F a s t a n d P r o v a b l y G o o d S e e d i n g s f o r k - M e a n s k-Means clustering F a s t a n d P r o v a b l y G o o d S e e d i n g s f o r k - M e a n s k-Means clustering Most popular clustering approach (nonconvex) F a s t a n d P r o v a b l y G o o d S e e d i n g s f o r k - M e a n s k-Means clustering Most popular clustering approach (nonconvex) SEEDING Find initial cluster centers F a s t a n d P r o v a b l y G o o d S e e d i n g s f o r k - M e a n s k-Means clustering Most popular clustering approach (nonconvex) SEEDING FINE-TUNING Find initial cluster centers Iteratively improve solution F a s t a n d P r o v a b l y G o o d S e e d i n g s f o r k - M e a n s k-Means clustering Most popular clustering approach (nonconvex) MANY LOCAL MINIMA MAY EXIST SEEDING FINE-TUNING Find initial cluster centers Iteratively improve solution F a s t a n d P r o v a b l y G o o d S e e d i n g s f o r k - M e a n s k-Means clustering Most popular clustering approach (nonconvex) MANY LOCAL MINIMA MAY EXIST SEEDING FINE-TUNING Find initial cluster centers Iteratively improve solution ENSURES THAT LOCAL MINIMUM IS REACHED