predict.amdai: Prediction method for 'amdai' class objects.

In adaptDA: Adaptive Mixture Discriminant Analysis

Description

The adaptive mixture discriminant analysis (AMDA) allows to adapt a model-based classifier to the situation a class represented in the test set may have not been encountered earlier in the learning phase.

Usage

## S3 method for class 'amdai'
predict(object, Y, K, maxit = 25, eps = 1e-05, ...)

Arguments

object	a supervised classifier genarated by the amdai function (an amdai object).
Y	the test data.
K	the number of expected classes in the test set.
maxit	the maximum number of iterations for the EM algorithm.
eps	the threshold value for the likelihood differences to stop the EM algorithm.
...	additional options for internal functions.

Value

A list is returned with the following fields:

model	the used model
mean	the updated mixture means
prop	the updated mixture proportions
var	the updated mixture covariance matrices
cls	the predicted labels for the test data
P	the posterior probabilities
crit	the criterion values for AIC, BIC and ICL

Author(s)

Charles Bouveyron

References

C. Bouveyron, Adaptive mixture discriminant analysis for supervised learning with unobserved classes, Journal of Classification, vol. 31(1), pp. 49-84, 2014.

Examples

set.seed(12345)

## Data simulation
data(iris)
Z.data = iris[,-5]
Z.cls = as.numeric(iris[,5])
Z.cls[as.numeric(iris[,5]==2)] = 3
Z.cls[as.numeric(iris[,5]==3)] = 2
N = 150

## Sampling
ind = sample(1:N,N)
X.data = Z.data[ind[1:(2*N/3)],]
X.cls = Z.cls[ind[1:(2*N/3)]]
X.data = X.data[X.cls!=3,]
X.cls = X.cls[X.cls!=3]
Y.data = Z.data[ind[(2*N/3+1):N],]
Y.cls = Z.cls[ind[(2*N/3+1):N]]

# Plotting the data
par(mfrow=c(2,3))#,cex.lab=0.75,cex.axis=0.75,cex.main=0.75,cex.sub=0.75)
pc = princomp(Z.data)
x = predict(pc,X.data)
y = predict(pc,Y.data)
plot(y,type='n',main='Learning data')
points(x[,1:2],col=X.cls+1,pch=19,main='Learning data')
y = predict(pc,Y.data)
plot(y[,1:2],col=1,pch=19,main='Test data')
plot(y[,1:2],col=Y.cls+1,pch=19,main='True labels of test data')

## Usual classification with QDA
c1 = qda(X.data,X.cls)
res1 = predict(c1,Y.data)
plot(y[,1:2],col=as.numeric(res1$class)+1,pch=19,main='QDA results')

## Classification with AMDAi
c2 = amdai(X.data,X.cls,model='qda')
B = rep(c(-Inf),5)
myPRMS <- vector(mode='list', length=7) # vector of lists!
for (i in 2:5){
  myPRMS[[i]] = predict(c2,Y.data,K=i)
	B[i] = myPRMS[[i]]$crit$bic
}
plot(2:5,B[2:5],type='b',xlab='Nb of components',ylab='AIC value',main='AIC values for AMDA')
res2 = myPRMS[[which.max(B)]]
plot(y[,1:2],col=res2$cls+1,pch=19,main='AMDA results')

## Classification results
cat("* Correct classification rates :\n")
cat("\tQDA:\t",sum(res1$class == Y.cls) / length(Y.cls),"\n")
print(table(res1$class,Y.cls))
cat("\tAMDAi:\t",sum(res2$cls == Y.cls) / length(Y.cls),"\n")
print(table(res2$cls,Y.cls))

adaptDA documentation built on May 30, 2017, 2:39 a.m.