A package for efficient computations of standard clustering comparison measures

Stable version on the CRAN.

install.packages("aricode")

The development version is available via:

devtools::install_github("jchiquet/aricode")

Computation of measures for clustering comparison (ARI, AMI, NID and even the (\chi^2) distance) are usually based on the contingency table. Traditional implementations (e.g., function adjustedRandIndex of package mclust) are in (\Omega(n + u v)) where

• (n) is the size of the vectors the classifications of which are to be compared,
• (u) and (v) are the respective number of classes in each vectors.

In aricode we propose an implementation, based on radix sort, that is in (\Theta(n)) in time and space.
Importantly, the complexity does not depends on (u) and (v). Our implementation of the ARI for instance is one or two order of magnitude faster than some standard implementation in R.

The functions included in aricode are:

• ARI: computes the adjusted rand index
• Chi2: computes the Chi-square statistics
• MARI/MARIraw: computes the modified adjusted rand index (Sundqvist et al, in preparation)
• NVI: computes the the normalized variation information
• NID: computes the normalized information distance
• NMI: computes the normalized mutual information
• AMI: computes the adjusted mutual information
• expected_MI: computes the expected mutual information
• entropy: computes the conditional and joint entropies
• clustComp: computes all clustering comparison measures at once

Here are some timings to compare the cost of computing the adjusted Rand Index with aricode or with the commonly used function adjustedRandIndex of the mclust package: the cost of the latter can be prohibitive for large vectors: