consensus_cluster: Consensus clustering

View source: R/consensus_cluster.R

consensus_clusterR Documentation

Consensus clustering


Runs consensus clustering across subsamples of the data, clustering algorithms, and cluster sizes.


  nk = 2:4,
  p.item = 0.8,
  reps = 1000,
  algorithms = NULL,
  nmf.method = c("brunet", "lee"),
  hc.method = "average",
  xdim = NULL,
  ydim = NULL,
  rlen = 200,
  alpha = c(0.05, 0.01),
  minPts = 5,
  distance = "euclidean", = c("none", "full", "sampled"),
  scale = TRUE,
  type = c("conventional", "robust", "tsne"),
  min.var = 1,
  progress = TRUE,
  seed.nmf = 123456, = 1, = NULL,
  time.saved = FALSE



data matrix with rows as samples and columns as variables


number of clusters (k) requested; can specify a single integer or a range of integers to compute multiple k


proportion of items to be used in subsampling within an algorithm


number of subsamples


vector of clustering algorithms for performing consensus clustering. Must be any number of the following: "nmf", "hc", "diana", "km", "pam", "ap", "sc", "gmm", "block", "som", "cmeans", "hdbscan". A custom clustering algorithm can be used.


specify NMF-based algorithms to run. By default the "brunet" and "lee" algorithms are called. See NMF::nmf() for details.


agglomeration method for hierarchical clustering. The the "average" method is used by default. Seestats::hclust() for details.


x dimension of the SOM grid


y dimension of the SOM grid


the number of times the complete data set will be presented to the SOM network.


SOM learning rate, a vector of two numbers indicating the amount of change. Default is to decline linearly from 0.05 to 0.01 over rlen updates. Not used for the batch algorithm.


minimum size of clusters for HDBSCAN. Default is 5.


a vector of distance functions. Defaults to "euclidean". Other options are given in stats::dist(). A custom distance function can be used.

Prepare the data on the "full" dataset, the "sampled" dataset, or "none" (default).


logical; should the data be centered and scaled?


if we use "conventional" measures (default), then the mean and standard deviation are used for centering and scaling, respectively. If "robust" measures are specified, the median and median absolute deviation (MAD) are used. Alternatively, we can apply "tsne" for dimension reduction.


minimum variability measure threshold used to filter the feature space for only highly variable features. Only features with a minimum variability measure across all samples greater than min.var will be used. If type = "conventional", the standard deviation is the measure used, and if type = "robust", the MAD is the measure used.


logical; should a progress bar be displayed?


random seed to use for NMF-based algorithms

seed to use to ensure each algorithm operates on the same set of subsamples

if not NULL, the returned array will be saved at each iteration as well as at the end of the function call to an rds object with as the file name.


logical; if TRUE, the date saved is appended to Only applicable when is not NULL.


See examples for how to use custom algorithms and distance functions. The default clustering algorithms provided are:

  • "nmf": Nonnegative Matrix Factorization (using Kullback-Leibler Divergence or Euclidean distance; See Note for specifications.)

  • "hc": Hierarchical Clustering

  • "diana": DIvisive ANAlysis Clustering

  • "km": K-Means Clustering

  • "pam": Partition Around Medoids

  • "ap": Affinity Propagation

  • "sc": Spectral Clustering using Radial-Basis kernel function

  • "gmm": Gaussian Mixture Model using Bayesian Information Criterion on EM algorithm

  • "block": Co-clustering using a multiple latent block model

  • "som": Self-Organizing Map (SOM) with Hierarchical Clustering

  • "cmeans": Fuzzy C-Means Clustering

  • "hdbscan": Hierarchical Density-based Spatial Clustering of Applications with Noise (HDBSCAN)

The progress bar increments on every unit of reps.


An array of dimension nrow(x) by reps by length(algorithms) by length(nk). Each cube of the array represents a different k. Each slice of a cube is a matrix showing consensus clustering results for algorithms. The matrices have a row for each sample, and a column for each subsample. Each entry represents a class membership.

When "hdbscan" is part of algorithms, we do not include its clustering array in the consensus result. Instead, we report two summary statistics as attributes: the proportion of outliers and the number of clusters.


The nmf.method options are "brunet" (Kullback-Leibler Divergence) and "lee" (Euclidean distance). When "hdbscan" is chosen as an algorithm to use, its results are excluded from the rest of the consensus clusters. This is because there is no guarantee that the cluster assignment will have every sample clustered; more often than not there will be noise points or outliers. In addition, the number of distinct clusters may not even be equal to nk.


Derek Chiu, Aline Talhouk


dat <- hgsc[1:100, 1:50]

# Custom distance function
manh <- function(x) {
  stats::dist(x, method = "manhattan")

# Custom clustering algorithm
agnes <- function(d, k) {
  return(as.integer(stats::cutree(cluster::agnes(d, diss = TRUE), k)))

assign("agnes", agnes, 1)

cc <- consensus_cluster(dat, reps = 6, algorithms = c("pam", "agnes"),
distance = c("euclidean", "manh"), progress = FALSE)

diceR documentation built on Sept. 29, 2023, 1:06 a.m.