Implement kmeans++ for better centroid initialization

okaneco / kmeans-colors

k-means clustering library and binary to find dominant colors in images

Apache License 2.0

133 stars 9 forks source link

Implement kmeans++ for better centroid initialization #20

Closed okaneco closed 4 years ago

okaneco commented 4 years ago

Change library kmeans algorithms to kmeans++ random initialization Change binary to use Hamerly for k greater than 1

k-means++ provides much faster results across the board and slightly less error in many cases. However, this changes the behavior from the previous implementation.