kernel_kmeans: Kernel k-means

In vittoriofortino84/COPS: Clustering algorithms for Omics-driven Patient Stratification

Kernel k-means

Description

Kernel k-means with different algorithm options. Spectral relaxation uses standard randomly initialized k-means on the eigen vectors of the kernel matrix while the QR decomposition of the eigen vectors yields a single solution directly. The last option is to use the kernel matrix to optimize average distances without utilizing the spectral relaxation.

Usage

kernel_kmeans(
  K,
  n_k,
  algorithm = "spectral_qr",
  spectral_qr_refine = TRUE,
  kernel_eigen_vectors = NULL,
  max_iter = 100,
  num_init = 100,
  tol = 1e-08,
  parallel = 1,
  ...
)

Arguments

K	kernel matrix
n_k	number of clusters
algorithm	one of "spectral", "spectral_qr", or "kernelized"
spectral_qr_refine	refine QR result with kernelized k-means
kernel_eigen_vectors	eigenvectors of the kernel matrix can be pre-computed
max_iter	maximum number of iterations
num_init	number of kmeans++ initializations for kernelized k-means and spectral clustering
tol	delta error convergence threshold for spectral clustering
parallel	number of threads for kernelized_kmeans
...	ignored

Details

The tol parameter is only used by the spectral relaxation algorithm which makes use of KMeans_rcpp. Other iterative algorithms are considered converged only if cluster assignments do not change.

Value

list of cluster assignments and k-means objective

vittoriofortino84/COPS documentation built on Jan. 28, 2025, 3:16 p.m.