CODEWITHSUNDEEP
rk-means
d-max
d-max

Reputation: 177

Creation prediction function for kmean in R

I want create predict function which predicts for which cluster, observation belong

data(iris)
 mydata=iris
m=mydata[1:4]
train=head(m,100)
xNew=head(m,10)


rownames(train)<-1:nrow(train)

norm_eucl=function(train)
  train/apply(train,1,function(x)sum(x^2)^.5)
m_norm=norm_eucl(train)


result=kmeans(m_norm,3,30)

predict.kmean <- function(cluster, newdata)
{
  simMat <- m_norm(rbind(cluster, newdata),
              sel=(1:nrow(newdata)) + nrow(cluster))[1:nrow(cluster), ]
  unname(apply(simMat, 2, which.max))
}

## assign new data samples to exemplars
predict.kmean(m_norm, x[result$cluster, ], xNew)

After i get the error

Error in predict.kmean(m_norm, x[result$cluster, ], xNew) : 
  unused argument (xNew)

i understand that i am making something wrong function, cause I'm just learning to do it, but I can't understand where exactly.

indeed i want adopt similar function of apcluster ( i had seen similar topic, but for apcluster)

predict.apcluster <- function(s, exemplars, newdata)
{
  simMat <- s(rbind(exemplars, newdata),
              sel=(1:nrow(newdata)) + nrow(exemplars))[1:nrow(exemplars), ]
  unname(apply(simMat, 2, which.max))
}

## assign new data samples to exemplars
predict.apcluster(negDistMat(r=2), x[apres@exemplars, ], xNew)

how to do it?

Upvotes: 1

Views: 1360

Answers (1)

Julius Vainora
Julius Vainora

Reputation: 48201

Rather than trying to replicate something, let's come up with our own function. For a given vector x, we want to assign a cluster using some prior k-means output. Given how k-means algorithm works, what we want is to find which cluster's center is closest to x. That can be done as

predict.kmeans <- function(x, newdata)
  apply(newdata, 1, function(r) which.min(colSums((t(x$centers) - r)^2)))

That is, we go over newdata row by row and compute the corresponding row's distance to each of the centers and find the minimal one. Then, e.g.,

head(predict(result, train / sqrt(rowSums(train^2))), 3)
# 1 2 3 
# 2 2 2
all.equal(predict(result, train / sqrt(rowSums(train^2))), result$cluster)
# [1] TRUE

which confirms that our predicting function assigned all the same clusters to the training observations. Then also

predict(result, xNew / sqrt(rowSums(xNew^2)))
#  1  2  3  4  5  6  7  8  9 10 
#  2  2  2  2  2  2  2  2  2  2

Notice also that I'm calling simply predict rather than predict.kmeans. That is because result is of class kmeans and a right method is automatically chosen. Also notice how I normalize the data in a vectorized manner, without using apply.

Upvotes: 6

Related Questions