fastpam {fastkmedoids}

FastPAM

Description

FastPAM: An improved version of PAM, that is usually O(k) times faster. Because of the speed benefits, we also suggest to use a linear-time initialization, such as the k-means++ initialization or the proposed LAB (linear approximative BUILD, the third component of FastPAM) initialization, and try multiple times if the runtime permits. (Schubert and Rousseeuw, 2019)

Usage

fastpam(
  rdist,
  n,
  k,
  maxiter = 0L,
  initializer = "LAB",
  fasttol = 1,
  seed = 123456789L
)

Arguments

rdist

The distance matrix (lower triangular matrix, column wise storage)

n

The number of observations

k

The number of clusters to produce.

maxiter

The maximum number of iterations (default: 0)

initializer

Initializer: either "BUILD" (used in classic PAM) or "LAB" (linear approximative BUILD) Because of the speed benefits, "LAB" is suggested, and one can try multiple times if the runtime permits.

fasttol

Tolerance for fast swapping behavior (may perform worse swaps). Default: 1.0, which means to perform any additional swap that gives an improvement. When set to 0, it will only execute an additional swap if it appears to be independent (i.e., the improvements resulting from the swap have not decreased when the first swap was executed).

seed

Seed for random number generator. Default: 123456789

Value

KMedoids S4 class

References

Erich Schubert, Peter J. Rousseeuw "Faster k-Medoids Clustering: Improving the PAM, CLARA, and CLARANS Algorithms" 2019 https://arxiv.org/abs/1810.05691