knn.probability: KNN prediction probability routine using pre-calculated...
In Dasonk/knnflex: A more flexible KNN

Description Usage Arguments Details Value Note Author(s) See Also Examples

K-Nearest Neighbor prediction probability method which uses the distances calculated by knn.dist. For predictions (not probabilities) see knn.predict.

1 2	knn.probability(train, test, y, dist.matrix, k = 1, ties.meth = "min")

`train`	indexes which specify the rows of x provided to `knn.dist` to be used in making the predictions
`test`	indexes which specify the rows of x provided to `knn.dist` to make predictions for
`y`	responses, see details below
`dist.matrix`	the output from a call to `knn.dist`
`k`	the number of nearest neighbors to consider
`ties.meth`	method to handle ties for the kth neighbor, the default is "min" which uses all ties, alternatives include "max" which uses none if there are ties for the k-th nearest neighbor, "random" which selects among the ties randomly and "first" which uses the ties in their order in the data

Prediction probabilities are calculated for each test case by aggregating the responses of the k-nearest neighbors among the training cases and using the classprob. k may be specified to be any positive integer less than the number of training cases, but is generally between 1 and 10.

The indexes for the training and test cases are in reference to the order of the entire data set as it was passed to knn.dist.

Only responses for the training cases are used. The responses provided in y may be those for the entire data set (test and training cases), or just for the training cases.

The ties are handled using the rank function. Further information may be found by examining the ties.method there.

a matirx of prediction probabilities whose number of columns is the number of test cases and the number of rows is the number of levels in the responses.

For the traditional scenario, classification using the Euclidean distance on a fixed set of training cases and a fixed set of test cases, the method knn is ideal. The functions knn.dist and knn.predict are intend to be used when something beyond the traditional case is desired. For example, prediction on a continuous y (non-classification), cross-validation for the selection of k, or the use of an alternate distance method are well handled.

Atina Dunlap Brooks

knn.dist, knn.predict, knn

# the iris example used by knn(class)
library(class)
data(iris3)
train <- rbind(iris3[1:25,,1], iris3[1:25,,2], iris3[1:25,,3])
test <- rbind(iris3[26:50,,1], iris3[26:50,,2], iris3[26:50,,3])
cl <- factor(c(rep("s",25), rep("c",25), rep("v",25)))
# how to get predictions from knn(class)
pred <- knn(train, test, cl, k = 3, prob=TRUE)
# display the confusion matrix
table(pred,cl)
# view probabilities (only the highest probability is returned)
attr(pred,"prob")

# how to get predictions with knn.dist and knn.predict
x <- rbind(train,test)
kdist <- knn.dist(x)
pred <- knn.predict(1:75, 76:150, cl, kdist, k=3)
# display the confusion matrix
table(pred,cl)
# view probabilities (all class probabilities are returned)
knn.probability(1:75, 76:150, cl, kdist, k=3)

# to compare probabilites, rounding done for display purposes
p1 <- knn(train, test, cl, k = 3, prob=TRUE)
p2 <- round(knn.probability(1:75, 76:150, cl, kdist, k=3), digits=2)
table( round(attr(p1,"prob"), digits=2), apply(p2,2,max) )

# note any small differences in predictions are a result of
# both methods breaking ties in majority class randomly