demo_scm.R

In aydindemircioglu/crushSVM: A Collection of SVM Ensemble Algorithms.

# 
	library(devtools)
	load_all(".")

	library(microbenchmark)
	library(e1071)
	library(BBmisc)
	library(mlr)
	
	# make sure things work from Rscript also
	library(methods)

# set Kpos, Kminus
	Kpos = 5
	Kneg = 10
	
# convert iris to matrix
	x = as.matrix(iris[,1:4])
	y = as.vector(as.numeric(iris[,5]))
	# make sure its binary
	y = replace(y, y == 2, 0)
	y = replace(y, y == 3, 0)

	posEx = as.data.frame(subset(x, y == 1))
	negEx = as.data.frame(subset(x, y == 0))
	
	clusterMethod = "KMeans"
	
	if (clusterMethod == "KMeans") {

	# cluster pos ex
		cluster.task = makeClusterTask(data = posEx)
		cluster.lrn = makeLearner("cluster.SimpleKMeans", N = Kpos)
		clusterModelPos = train(cluster.lrn, task = cluster.task)
		clusteringPos = predict(clusterModelPos, newdata=posEx)

	# cluster neg ex
		cluster.task = makeClusterTask(data = negEx)
		cluster.lrn = makeLearner("cluster.SimpleKMeans", N = Kneg)
		clusterModelNeg = train(cluster.lrn, task = cluster.task)
		clusteringNeg = predict(clusterModelNeg, newdata=negEx)
	} 
	
	if (clusterMethod == "EM") {
	# now for each do X
		#clusteringNeg$data
		
		# em clustering in mlr
	#	cluster.task = makeClusterTask(data = negEx)
		#cluster.lrn = makeLearner("cluster.EM", N = Kneg)

	}

	# compute the kernel
	K = ..
	
	# compute the weights, this depends on the clustering
	weights = ...
	
	# call SVM
	sharksvm (x, y, weights = weights, type = "CSVM", kernel = "precomputed", kernelmatrix = K, verbose = TRUE, C = 1, gamma = 1)