| SOM Toolbox | Online documentation | http://www.cis.hut.fi/projects/somtoolbox/ |
[C,P]=knn(d, Cp, K)
KNN K-Nearest Neighbor classifier using an arbitrary distance matrix
[C,P]=knn(d, Cp, [K])
Input and output arguments ([]'s are optional):
d (matrix) of size NxP: This is a precalculated dissimilarity (distance matrix).
P is the number of prototype vectors and N is the number of data vectors
That is, d(i,j) is the distance between data item i and prototype j.
Cp (vector) of size Px1 that contains integer class labels. Cp(j) is the class of
jth prototype.
[K] (scalar) the maximum K in K-NN classifier, default is 1
C (matrix) of size NxK: integers indicating the class
decision for data items according to the K-NN rule for each K.
C(i,K) is the classification for data item i using the K-NN rule
P (matrix) of size NxkxK: the relative amount of prototypes of
each class among the K closest prototypes for each classifiee.
That is, P(i,j,K) is the relative amount of prototypes of class j
among K nearest prototypes for data item i.
If there is a tie between representatives of two or more classes
among the K closest neighbors to the classifiee, the class i selected randomly
among these candidates.
IMPORTANT If K>1 this function uses 'sort' which is considerably slower than
'max' which is used for K=1. If K>1 the knn always calculates
results for all K-NN models from 1-NN up to K-NN.
EXAMPLE 1
sP; % a SOM Toolbox data struct containing labeled prototype vectors
[Cp,label]=som_label2num(sP); % get integer class labels for prototype vectors
sD; % a SOM Toolbox data struct containing vectors to be classified
d=som_eucdist2(sD,sP); % calculate euclidean distance matrix
class=knn(d,Cp,10); % classify using 1,2,...,10-rules
class(:,5); % includes results for 5NN
label(class(:,5)) % original class labels for 5NN
EXAMPLE 2 (leave-one-out-crossvalidate KNN for selection of proper K)
P; % a data matrix of prototype vectors (rows)
Cp; % column vector of integer class labels for vectors in P
d=som_eucdist2(P,P); % calculate euclidean distance matrix PxP
d(eye(size(d))==1)=NaN; % set self-dissimilarity to NaN:
% this drops the prototype itself away from its neighborhood
% leave-one-out-crossvalidation (LOOCV)
class=knn(d,Cp,size(P,1)); % classify using all possible K
% calculate and plot LOOC-validated errors for all K
failratep = ...
100*sum((class~=repmat(Cp,1,size(P,1))))./size(P,1); plot(1:size(P,1),failratep)