ssClust: ssClust

In Noobivsho/ssClust: GMM for semi-supervised case a-la Mclust

Description

Performs GMM using some supervision

Usage

ssClust(X, knownLabels = NULL, trueLabels = NULL,
                 knownCannotLink = NULL, cannotLinkWithIdx = NULL,
                 Grange = 2, modelNames = c("VVV"), runParallel = TRUE,
                 fracOfCores2Use = 1, initClassAssignments = NULL,
                 initializationStrategy = "kpp",
         penalizeSupervised=T)

Arguments

X	data
knownLabels	vector of indices of rows of x whose labels are known
trueLabels	length nrow(x) vector of true labels (only trueLabels[knownLabels] matter)
knownCannotLink	vector of indices of data with cannot link with constraints
cannotLinkWithIdx	cannotLinkWithIdx[[“i"]] = data i cannot link with these indices (in vector) ex: Suppose cannotLinkWithIdx[[“2"]]=c(1,3). Then, 2 cannot link with 1 and 3. Constructed with, for example: knownCannotLink = c(3,4,5,6,7,8) cannotLinkWithIdx = new.env() for(i in knownCannotLink) cannotLinkWithIdx[[as.character(i)]]=c(1,2)
Grange	number of clsuters considered
modelNames	models considered
runParallel	boolean for whether to run in parallel
fracOfCores2Use	Fraction of all cores to use (if in parallel)
initClassAssignments	Hard initial class assignment which overrides the default hiearchical clustering initialization scheme.
initializationStrategy	Strategy to initialize cluster labels for EM algorithm. Currently only supports kpp for semi-supervised k-means++
penalizeSupervised	Boolean for whether or not to include supervised data in the penalty term in the BIC. Defaults to TRUE.

Value

selectedModel	Model selected (num components and covariance constraints)
BIC	Bayesian Information Criteria values for all models, adjusted for alternative penalty (dont use this number unless you know what you're doing))
BICunadjusted	Traditional Bayesian Information Criteria for all models
allModels	BIC and likelihood for all models
z	posterior probabilities of class memberships
parameters	winning model's parameters, like Mclust's parameters. Item:pro–mixing parameters Item: mean–means Item: variance–variance of the form mclustVariance
classes	vector of class assignments for winning model
numParams	number of parameters estimated for all models
loglik	loglik of winning model
initPart	initial partition
modelsClasses	matrix of class assignments for all models considered (generally ignore this)

Author(s)

Jordan Yoder

References

http://www.stat.washington.edu/mclust/

Examples

library('MASS')
library('mclust')

########### GENERATE DATA#######
n = 500 #total number of data points
C = 3 #total number of clusters
c = 5 #total number of labeled data point
meanOfControlCluster=-1
varOfControlCluster = 1/5
trueLabels = sort(sample(1:C,n,replace=TRUE)-1) #equal prior proportions

#generate the data
tableLabel = table(trueLabels)
cumSumTableLabel = cumsum(tableLabel)
cumSumTableLabel = c(0,cumSumTableLabel)
dataDim = 5
X = matrix(0, n,dataDim)
#save the true parameters
trueMeans = NULL
trueCovMat = NULL
spreadFactor=5 
trueMeans[[1]]=rep(meanOfControlCluster,dataDim)
trueMeans[[2]]=rep(1,dataDim)
trueMeans[[3]]=rep(-1,dataDim)
trueCovMat[[1]]=diag(varOfControlCluster,dataDim)
trueCovMat[[2]]=diag(1,dataDim)
trueCovMat[[3]]=diag(1,dataDim)
for(i in 1:C){
  X[(cumSumTableLabel[i]+1):(cumSumTableLabel[i+1]),] = 
  mvrnorm(n=tableLabel[i],mu=trueMeans[[i]],Sigma=trueCovMat[[i]])
}

##### CLUSTER USING ssClust and Mclust ######
mclust.out <- Mclust(X,G = 2:5, modelNames = c('VVV','EEE','VVI','VII') )
ssClust.out<-ssClust(X,knownLabels=1:c,trueLabels=trueLabels,Grange=2:5,
modelNames=c('VVV','EEE','VVI','VII'), runParallel=FALSE)
retVal <- c(adjustedRandIndex(mclust.out$cl[-(1:c)],trueLabels[-(1:c)]),
            adjustedRandIndex(ssClust.out$cl[-(1:c)],trueLabels[-(1:c)]))
retVal <- ifelse(is.nan(retVal),-1,retVal)

Noobivsho/ssClust documentation built on Aug. 10, 2019, 5:47 a.m.