ClusterR - optimal_init method

Question

The documentation of the initialization methods for clustering using ClusterR is very enigmatic. Are you able to post a reference to a paper describing optimal_init method?

This is what is stated officially:

optimal_init: this initializer adds rows of the data incrementally, while checking that they do not already exist in the centroid-matrix

Answer 1

The initializers used in the KMeans_rcpp and MiniBatchKmeans of the ClusterR package are :

• random (random selection of data rows as initial centroids),
• kmeans++ (Reference : http://theory.stanford.edu/~sergei/papers/kMeansPP-soda.pdf AND http://stackoverflow.com/questions/5466323/how-exactly-does-k-means-work),
• quantile_init (it uses quantiles and cumulative distances to pick the initial centroids) and
• optimal_init (it adds observations to the centroids incrementally by first checking that they do not already exist in the picked centroids).

I added the last two (quantile_init and optimal_init) in the package, because I found out after testing in various data sets that they give similar (or better) results (using validation metrics) and/or run faster. They are both experimental and (you are right) in the next version of the package I'll add a note in the documentation details. You can see the rcpp code of quantile_init and optimal_init in the package repository.