CLARA
A sampling-based method that scales PAM/k-medoids clustering by optimizing medoids on several subsets rather than the full dataset.
CLARA does not change the k-medoids objective; it reduces the cost by running PAM on several smaller samples and evaluating the resulting medoid candidates against the full dataset.
When It Fits
It is useful when building or retaining a full dissimilarity structure is too expensive. Because medoids are actual observations, the resulting cluster representatives can also be easier to inspect than synthetic centroids.
Limitation
A structure absent from the sampled subsets cannot be recovered by the optimizer. Stability across samples, cluster-size distribution, and the final dissimilarity should therefore be inspected together.
Source
- https://stat.ethz.ch/R-manual/R-devel/library/cluster/html/clara.html