Strategy for initializing KMeans centroids.
random: Centroids are sampled randomly from the data. This has complexity
kmeans++: Centroids are computed iteratively. The first centroid is sampled randomly from the data. Subsequently, centroids are sampled from the remaining datapoints with probability proportional to
D(x)is the distance of datapoint
xto the closest centroid chosen so far. This has complexity
ndatapoints. If done on mini-batches, the complexity increases to