Silhouette of NMF Clustering

Usage

S3 (NMF)
`silhouette`(x, what = NULL, order = NULL, ...)

Arguments

x

an NMF object, as returned by nmf.

what

defines the type of clustering the computed silhouettes are meant to assess: 'samples' for the clustering of samples (i.e. the columns of the target matrix), 'features' for the clustering of features (i.e. the rows of the target matrix), and 'chc' for the consensus clustering of samples as defined by hierarchical clustering dendrogram, 'consensus' for the consensus clustering of samples, with clustered ordered as in the default hierarchical clustering used by consensusmap when plotting the heatmap of the consensus matrix (for multi-run NMF fits). That is dist = 1 - consensus(x), average linkage and reordering based on row means.

order

integer indexing vector that can be used to force the silhouette order.

...

extra arguments not used.

Examples

x <- rmatrix(100, 20, dimnames = list(paste0('a', 1:100), letters[1:20]))
res <- nmf(x, 4, nrun = 5)

# sample clustering from best fit
plot(silhouette(res))

# from consensus
plot(silhouette(res, what = 'consensus'))

# feature clustering
plot(silhouette(res, what = 'features'))

# average silhouette are computed in summary measures
summary(res)

##   Length    Class     Mode 
##        1 NMFfitX1       S4

# consensus silhouettes are ordered as on default consensusmap heatmap
op <- par(mfrow = c(1,2))
consensusmap(res)
si <- silhouette(res, what = 'consensus')
plot(si)

par(op)

# if the order is based on some custom numeric weights
op <- par(mfrow = c(1,2))
cm <- consensusmap(res, Rowv = runif(ncol(res)))
# NB: use reverse order because silhouettes are plotted top-down
si <- silhouette(res, what = 'consensus', order = rev(cm$rowInd))
plot(si)

par(op)

# do the reverse: order the heatmap as a set of silhouettes
si <- silhouette(res, what = 'features')
op <- par(mfrow = c(1,2))
basismap(res, Rowv = si)
plot(si)

par(op)

See also

predict