Cluster Analysis using diceR


Cluster analysis is a way of “slicing and dicing” data to allow the grouping together of similar entities and the separation of dissimilar ones. Issues arise due to the existence of a diverse number of clustering algorithms, each with different techniques and inputs, and with no universally optimal methodology. Thus, a framework for cluster analysis and validation methods are needed. Our approach is to use cluster ensembles from a diverse set of algorithms so that the final class labels. This ensures that the data has been considered from several angles and using a variety of methods. We have currently implemented about 15 clustering algorithms, and we provide a simple framework to add additional algorithms (see example("consensus_cluster")). Although results are dependent on the subset of algorithms chosen for the ensemble, the intent is that we capture a variety of clustering and select those that are most consistent with the data.

You can install diceR from CRAN with:


Or get the latest development version from GitHub:

# install.packages("devtools")
hgsc <- hgsc[1:100, 1:50]

We load an excerpt of an expression data set from TCGA of 100 high grade serous carcinoma samples measured on 50 genes.

Clustering Algorithms

The list of clustering algorithms available for use are:

They are passed as a character vector into the algorithms parameter of consensus_cluster(). Note that the list continues to evolve as we implement more algorithms into diceR.

Cluster Models

The clustering algorithms can be categorized under different models or paradigms based on how the clusters are formed. We provide a brief overview to guide the initial selection of algorithms since no single algorithm works for every data model.1

Distance Measures

The list of distance measures available for use are those found in stats::dist and the spearman distance:

They are passed as a character vector into the distance parameter of consensus_cluster(). Only the "hc", "diana", "km", "pam" algorithms use the distance parameter.

Validity Indices


The list of internal validity indices used for evaluating clusters are:


The list of external validity indices used for evaluating clusters are:

Consensus Clustering

When Monti et al. (2003) first introduced consensus clustering, the idea was to use one clustering algorithm on bootstrapped subsamples of the data. We implement some extensions where a consensus is reached across subsamples and across algorithms. The final cluster assignment is then computed using statistical transformations on the ensemble cluster.

The base function of this package is consensus_cluster(), which outputs cluster assignments across subsamples and algorithms, for different k (number of clusters). For example, let’s say we were interested in clustering the hgsc data into 3 or 4 clusters, using 80% resampling on 5 replicates, for these clustering algorithms: Hierarchical Clustering, PAM, and DIvisive ANAlysis Clustering (DIANA). Euclidean distance is used for all algorithms.

CC <- consensus_cluster(hgsc, nk = 3:4, p.item = 0.8, reps = 5,
                        algorithms = c("hc", "pam", "diana"))

The output is a 4-dimensional array: rows are samples, columns are different bootstrap subsample replicates, slices are algorithms, and each “box” (4th dimension) is for a different k. Below are the first few cluster assignments for each replicate in the DIANA algorithm for k = 3.

co <- capture.output(str(CC))
strwrap(co, width = 80)
#> [1] "int [1:100, 1:5, 1:3, 1:2] 1 3 1 3 1 1 1 NA 1 1 ..."                   
#> [2] "- attr(*, \"dimnames\")=List of 4"                                     
#> [3] "..$ : chr [1:100] \"TCGA.04.1331_PRO.C5\" \"TCGA.04.1332_MES.C1\""     
#> [4] "\"TCGA.04.1336_DIF.C4\" \"TCGA.04.1337_MES.C1\" ..."                   
#> [5] "..$ : chr [1:5] \"R1\" \"R2\" \"R3\" \"R4\" ..."                       
#> [6] "..$ : chr [1:3] \"HC_Euclidean\" \"PAM_Euclidean\" \"DIANA_Euclidean\""
#> [7] "..$ : chr [1:2] \"3\" \"4\""
Cluster Assignments for DIANA, k = 3
  R1 R2 R3 R4 R5
TCGA.04.1331_PRO.C5 1 2 NA NA NA
TCGA.04.1332_MES.C1 1 2 1 1 NA
TCGA.04.1336_DIF.C4 2 NA NA 1 NA
TCGA.04.1337_MES.C1 2 1 3 1 2
TCGA.04.1338_MES.C1 2 1 3 1 2
TCGA.04.1341_PRO.C5 1 2 1 2 1

Note the unavoidable presence of NAs because we used 80% subsampling. This can be problematic in certain downstream ensemble methods, so we can impute missing values using K-Nearest Neighbours beforehand. There might still be NAs after KNN because of how the decision threshold was set. The remaining missing values can be imputed using majority voting.

CC <- apply(CC, 2:4, impute_knn, data = hgsc, seed = 1)
CC_imputed <- impute_missing(CC, hgsc, nk = 4)
#> [1] 5
#> [1] 0

We can carry on the analysis using either CC or CC_imputed.

Consensus Functions

diceR provides functions for retrieving useful summaries and other results for consensus clustering.

Compute consensus matrix with consensus_matrix()

The consensus matrix is an n by n matrix, where n is the number of samples. Each element is a real-valued number between 0 and 1 inclusive, representing the proportion of times that two samples were clustered together out of the times that the same samples were chosen in the bootstrap resampling. The diagonal are all one’s. Suppose we wanted to compute the consensus matrix for PAM, k = 4, and visualize using graph_heatmap():

Combine consensus summaries with consensus_combine()

If we wish to separately extract consensus matrices and consensus classes for every algorithm, consensus_combine() is a convenient wrapper to do so. Setting element = "matrix" returns a list of consensus matrices. On the other hand, setting element = "class" returns a matrix with rows as samples, and columns as clustering assignments for each algorithm.

Consensus Classes
HC_Euclidean PAM_Euclidean DIANA_Euclidean
1 1 1
2 1 1
1 2 2
3 1 2
1 2 2
1 1 1

One can feed in ccomb_class (instead of CC) into consensus_matrix() to obtain a consensus matrix across subsamples and algorithms (and k).

A situation might also arise where we initially decided on using 3 clustering algorithms for the ensemble, but later wish to add additional algorithms for analysis. consensus_combine() takes in any number of ensemble objects (e.g. CC and CC2) and combines the results.

Consensus Classes with KM added
HC_Euclidean PAM_Euclidean DIANA_Euclidean KM_Euclidean
1 1 1 1
2 1 1 1
1 2 2 2
3 1 2 3
1 2 2 2
1 1 1 1

Evaluate, trim, and reweigh algorithms with consensus_evaluate()

Internal cluster validation indices assess the performance of results by taking into account the compactness and separability of the clusters. We choose a variety of indices on which to compare the collection of clustering algorithms. We use the PAC (Proportion of Ambiguous Clusters), the proportion of entries in a consensus matrix that are strictly between lower (defaults to 0) and upper (defaults to 1), to give a measure of cluster stability. In addition, if no reference class is provided, we calculate the average PAC across algorithms within each k, and choose the k with the greatest average PAC. If there is a reference class, k is the number of distinct classes in the reference.

Internal Indices for k = 4 (continued below)
Algorithms calinski_harabasz dunn pbm tau gamma c_index davies_bouldin
HC_Euclidean 3.949 0.4524 13.25 0.3188 0.7036 0.2092 0.5132
PAM_Euclidean 15.09 0.2979 8.456 0.3756 0.5701 0.1796 2.056
DIANA_Euclidean 12.51 0.2663 5.698 0.3863 0.5516 0.1909 1.507
KM_Euclidean 15.84 0.2541 8.839 0.4683 0.7021 0.1195 1.912
mcclain_rao sd_dis ray_turi g_plus silhouette s_dbw Compactness Connectivity
0.7278 0.2666 0.5899 0.03044 NA NA 7.783 15.93
0.782 0.4754 1.225 0.09328 0.08144 NA 6.875 72.87
0.7848 0.4673 1.685 0.11 0.02748 NA 7.071 50.82
0.7398 0.4449 1.142 0.06624 0.06399 NA 6.84 73.99

We see that the biclustering algorithm is the least ambiguous and also most well-clustered (high compactness and separability).

Some algorithms perform too poorly to deserve membership in the cluster ensemble. We consider the relative ranks of each algorithm across all internal indices, and compute their sum. All algorithms below a certain quantile for the sum rank are trimmed (removed). By default this quantile is 75%.

After trimming, we can optionally choose to reweigh the algorithms based on the internal index magnitudes. Of course, we take into account the direction of optimality (higher is better is lower is better). Algorithms reweighed are then fed into the consensus functions. This is done by replicating each algorithm by a scalar factor that is proportional to its weight. For example, if we have two algorithms A and B, and A is given a weight of 80% and B is given a weight of 20%, then we make 4 copies of A and 1 copy of B. To minimize computational time, the total number of copies out of all algorithms has an upper bound of 100. Without reweighing, each algorithm is given equal weight.

The return value shows which algorithms were kept, removed (if any), and the trimmed (and potentially reweighed) cluster ensemble.

Significance Testing

To test whether there are four statistically distinct clusters (k = 4) versus no clusters (k = 1) using the PAM algorithm, we run sigclust with 50 simulations and generate a p-value for this significance test.

pam_4 <- ccomb_class2$`4`[, "PAM_Euclidean"]
sig_obj <- sigclust(hgsc, k = 4, nsim = 100, labflag = 0, label = pam_4)
co <- capture.output(str(sig_obj))
strwrap(co, width = 80)
#>  [1] "Formal class 'sigclust' [package \"sigclust\"] with 10 slots"               
#>  [2] "..@ : num [1:100, 1:50] -0.0107 -0.7107 0.8815 -1.0851 -0.9322 ..."
#>  [3] ".. ..- attr(*, \"dimnames\")=List of 2"                                     
#>  [4] ".. .. ..$ : chr [1:100] \"TCGA.04.1331_PRO.C5\" \"TCGA.04.1332_MES.C1\""    
#>  [5] "\"TCGA.04.1336_DIF.C4\" \"TCGA.04.1337_MES.C1\" ..."                        
#>  [6] ".. .. ..$ : chr [1:50] \"ABAT\" \"ABHD2\" \"ACTB\" \"ACTR2\" ..."           
#>  [7] "..@ veigval : num [1:50] 11.81 4.51 2.66 2.29 1.84 ..."                     
#>  [8] "..@ vsimeigval: num [1:50] 11.81 4.51 2.66 2.29 1.84 ..."                   
#>  [9] "..@ simbackvar: num 0.42"                                                   
#> [10] "..@ icovest : num 2"                                                        
#> [11] "..@ nsim : num 100"                                                         
#> [12] "..@ simcindex : num [1:100] 0.67 0.634 0.641 0.637 0.606 ..."               
#> [13] "..@ pval : num 0.14"                                                        
#> [14] "..@ pvalnorm : num 0.163"                                                   
#> [15] "..@ xcindex : num 0.63"

The p-value is 0.14, indicating we do not have sufficient evidence to conclude there are four distinct clusters. Note that we did not use the full hgsc data set in this example, so the underlying biological mechanisms may not be fully captured.


  1. Estivill-Castro, Vladimir (20 June 2002). “Why so many clustering algorithms — A Position Paper”. ACM SIGKDD Explorations Newsletter. 4 (1): 65–75.
  2. Brendan J. Frey; Delbert Dueck (2007). “Clustering by passing messages between data points”. Science. 315 (5814): 972–976.
  3. C. Fraley, A. E. Raftery, T. B. Murphy and L. Scrucca (2012). mclust Version 4 for R: Normal Mixture Modeling for Model-Based Clustering, Classification, and Density Estimation. Technical Report No. 597, Department of Statistics, University of Washington.
  4. Campello, R. J. G. B.; Moulavi, D.; Sander, J. (2013). Density-Based Clustering Based on Hierarchical Density Estimates. Proceedings of the 17th Pacific-Asia Conference on Knowledge Discovery in Databases, PAKDD 2013, Lecture Notes in Computer Science 7819, p. 160.
  5. Campello, Ricardo JGB, et al. “Hierarchical density estimates for data clustering, visualization, and outlier detection.” ACM Transactions on Knowledge Discovery from Data (TKDD) 10.1 (2015): 5.
  6. G. Govaert and M. Nadif. Latent block model for contingency table. Communications in Statistics - Theory and Methods, 39(3):416–425, 2010.
  7. Isa, Dino, V. P. Kallimani, and Lam Hong Lee. “Using the self organizing map for clustering of text documents.” Expert Systems with Applications 36.5 (2009): 9584-9591.
  8. Andrew Y. Ng, Michael I. Jordan, Yair Weiss On Spectral Clustering: Analysis and an Algorithm Neural Information Processing Symposium 2001
  9. Brunet J, Tamayo P, Golub TR and Mesirov JP (2004). “Metagenes and molecular pattern discovery using matrix factorization.” Proceedings of the National Academy of Sciences of the United States of America, 101(12), pp. 4164-9.