knn.predict {knnflex} | R Documentation |
K-Nearest Neighbor prediction method which uses the distances calculated by
knn.dist
.
knn.predict(train, test, y, dist.matrix, k = 1, agg.meth = if (is.factor(y)) "majority" else "mean", 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 |
agg.meth |
method to combine responses of the nearest neighbors, defaults to "majority" for classification and "mean" for continuous responses |
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 |
Predictions are calculated for each test case by aggregating the responses of the
k-nearest neighbors among the training cases. 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 aggregation may be any named function. By default, classification (factored responses) will use the "majority" class function and non-factored responses will use "mean". Other options to consider include "min", "max" and "median".
The ties are handled using the rank
function. Further information may be found
by examining the ties.method
there.
a vector of predictions whose length is the number of test cases.
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
# a quick classification example x1 <- c(rnorm(20,mean=1),rnorm(20,mean=5)) x2 <- c(rnorm(20,mean=5),rnorm(20,mean=1)) x <- cbind(x1,x2) y <- c(rep(1,20),rep(0,20)) train <- sample(1:40,30) # plot the training cases plot(x1[train],x2[train],col=y[train]+1,xlab="x1",ylab="x2") # predict the other cases test <- (1:40)[-train] kdist <- knn.dist(x) preds <- knn.predict(train,test,y,kdist,k=3,agg.meth="majority") # add the predictions to the plot points(x1[test],x2[test],col=as.integer(preds)+1,pch="+") # display the confusion matrix table(y[test],preds) # 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) # display the confusion matrix table(pred,cl) # 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) # note any small differences are a result of both methods # breaking ties in majority class randomly # 5-fold cross-validation to select k for above example fold <- sample(1:5,75,replace=TRUE) cvpred <- matrix(NA,nrow=75,ncol=10) for (k in 1:10) for (i in 1:5) cvpred[which(fold==i),k] <- knn.predict(train=which(fold!=i),test=which(fold==i),cl,kdist,k=k) # display misclassification rates for k=1:10 apply(cvpred,2,function(x) sum(cl!=x))