partitionMap: Partition Maps

In partitionMap: Partition Maps

Description

using Random Forest multiclass output, embed observations in low-dimensional space

Usage

partitionMap(X, Y, XTEST = NULL, YTEST = NULL,  method = "pm", dimen = 2, 
                   force = TRUE, ntree = 100,
                   plottrain = TRUE, addjitter = 0.03, ...)

Arguments

X	matrix with predictor variables in the training dataset
Y	response variable, a factor with multiple classes
XTEST	The matrix of predictor variables for the test dataset (optional)
YTEST	Class labels of test observations, used for coloring the test embeddings in the plot. If not supplied, test observations are shwon in grey (optional)
method	pm for "partitonMap" and ha for "Homogeneity Analysis"
dimen	dimension of embedding, typically 2 or 3
force	use force-based variation of "partitionMap" algorithm? no effect if method="ha"
ntree	number of trees to use for randomForest prediction
plottrain	plot embedding for training data?
addjitter	amount if jitter to add to the plots to avoid overlapping observations (set addjitter=0 for no jitter)
...	other arguments to be passed to randomForest

Value

A list with values

Samples	low-dimensional co-ordinates of embedded training samples
Rules	low-dimensional co-ordinates of embedded Rules (nodes in the trees)
Z	a binary matrix, with as many rows as training samples and as many columns as rules. a value 1 in row i and column j indicates that observation i is part of rule j
Samplestest	low-dimensional co-ordinates of embedded test samples
Ztest	a binary matrix, with as many rows as test samples and as many columns as rules. a value 1 in row i and column j indicates that observation i in the test data is part of rule j
rf	the trained Random Forest classifier

Author(s)

Nicolai Meinshausen <[e-mail]>

References

Nicolai Meinshausen (2011)

Partition Maps

JCGS 20(4), 1007-1028

Examples

	
##---- load Soybean data ----
	data(Soybean)
	X <- Soybean[,-1]
	Y <- Soybean$Y 
	
##---- divide into training and test data ----
	indtrain <- rep(0,nrow(X))
	indtrain[sample(1:length(indtrain), ceiling(nrow(X)/3*2))] <- 1
	XTEST <- X[indtrain==0,]
	YTEST <- Y[indtrain==0]
	X <- X[indtrain==1,]
	Y <- Y[indtrain==1]


##---- compute Partition Map solution ----
	pm <- partitionMap(X,Y,XTEST=XTEST,method="pm",force=TRUE,
                                dimen=2,ntree=80,plottrain=TRUE)


##---- plot the embedded training and test samples ----
	par(mfrow=c(1,3))
	plot(pm$Samples,col=Y,pch=20,cex=1.5,main="Training Data",
                                    xlab="Dimension 1",ylab="Dimension 2")
	points(pm$Rules,pch=".")
	plot(pm$Samplestest,col=YTEST,pch=20,cex=1.5,main="Test Data",
                                     xlab="Dimension 1",ylab="Dimension 2")
	points(pm$Rules,pch=".")
	plot(pm$Samples,col=Y,pch=20,cex=1.5,xlab="",ylab="",type="n",axes=FALSE)
	legend(quantile(pm$Samples[,1],0),quantile(pm$Samples[,2],1),unique(Y),
                              col=1:length(unique(Y)),fill=1:length(unique(Y)),border=0)
	par(mfrow=c(1,1))

partitionMap documentation built on May 2, 2019, 2:43 a.m.