Maximal Information-Based Nonparametric Exploration for Variable Analysis

Wrapper for 'minepy' implementation of Maximal Information-based Nonparametric Exploration statistics (MIC and MINE family). Detailed information of the ANSI C implementation of 'minepy' can be found at < http://minepy.readthedocs.io/en/latest>.

minerva

R package for Maximal Information-Based Nonparametric Exploration computation

• Minepy Homepage
• Minepy Github
• Mictools Github

Install

• Latest cran release

install.packages("minerva")

• Development version

devtools::install_github('filosi/minerva')

Usage

• Basic usage with helper function mine.

library(minerva)

x <- 0:200 / 200
y <- sin(10 * pi * x) + x
mine(x,y, n.cores=1)

• Compute a single measure from the MINE suite using mine_stat.
  • Available mesures are: mic, mas, mev, mcn, tic, gmic

x <- 0:200 / 200
y <- sin(10 * pi * x) + x
mine_stat(x, y, measure="mic")

• To compute the mic-r2 measure use the cor R function:

x <- 0:200 / 200
y <- sin(10 * pi * x) + x

r2 <- cor(x, y)
mm <- mine_stat(x, y, measure="mic")
mm - r2**2

## mine(x, y, n.cores=1)[[5]]

Compute statistic on matrices

• All features in a single matrix (mine_compute_pstat).
• All possible combination of features between two matrices (mine_compute_cstat).
  • When comparing two matrices the function check for euquality of number of rows of the two matrices. If the matrices have different number of rows then an error is thrown.

x <- matrix(rnorm(1000), ncol=10, nrow=10)
y <- as.matrix(rnorm(1000), ncol=10, nrow=20)

## Compare feature of the same matrix
pstats(x)

## Compare features of matrix x with feature in matrix y
cstats(x, y)

Mictools pipeline

This is inspired to the original implementation by Albanese et al. available in python here: https://github.com/minepy/mictools.

Reading the data from mictool repository

datasaurus <- read.table("https://raw.githubusercontent.com/minepy/mictools/master/examples/datasaurus.txt", 
header=TRUE, row.names=1, as.is=TRUE, stringsAsFactors=FALSE)
datasaurus.m <- t(datasaurus)

Compute null distribution for tic_e

Automatically compute:

• tic_e null distribution based on permutations.
• histogram of the distribution with cumulative distribution.
• Observed values of tic_e for each pair of variable in datasaurus.
• Observed distribution of tic_e.
• P-value for each variable pair association.

ticnull <- mictools(datasaurus.m, nperm=10000, seed=1234)

## Get the names of the named list
names(ticnull)
##[1]  "tic"      "nulldist" "obstic"   "obsdist"  "pval"

Null Distribution

ticnull$nulldist

Observed distribution

ticnull$obsdist

Plot tic_e and pvalue distribution.

hist(ticnull$tic)

hist(ticenull$pval, breaks=50, freq=FALSE)

Use p.adjust.method to use a different pvalue correction method, or use the qvalue package to use Storey's qvalue.

## Correct pvalues using qvalue
qobj <- qvalue(ticnull$pval$pval)

## Add column in the pval data.frame
ticnull$pval$qvalue <- qobj$qvalue
ticnull$pval

Same table as above with the qvalue column added at the end.

Strenght of the association (MIC)

## Use columns of indexes and FDR adjusted pvalue 
micres <- mic_strength(datasaurus.m, ticnull$pval, pval.col=c(6, 2, 3))

Association strength computed based on the qvalue adjusted pvalue

## Use qvalue adjusted pvalue 
micresq <- mic_strength(datasaurus.m, ticnull$pval, pval.col=c("qvalue", "Var1", "Var2"))

Citing minepy/minerva and mictools