% Tony Hyun Kim
% CS 229, PS#3, Problem 6
% Implementation of k-means clustering
%------------------------------------------------------------
% Format:
%   examples (n by m):
%       examples(:,i) represents the i-th example
%   us (n by k):
%       us(:,i) represents the i-th mean
%------------------------------------------------------------
function [us J iter] = my_kmeans(examples,k)

[~, m] = size(examples);

% Initialize us to k-randomly chosen examples
p = randperm(m);
p = p(1:k);
us = examples(:,p);

% Vector that keeps track of which centroid has
%   been assigned to each example
cs = zeros(1,m);

% Our criterion for convergence will be that the
%   distortion function J changes by less than tol
%   OR we reach a maximum number of iterations
iter = 0;
maxIter = 1000;
tol  = 1e-6;
Jold = 1;
J = 0;
% We continue the optimization as long as the distortion function
%   J is improved by more than tol. As per the instructions, we
%   also enforce a minimum of 30 iterations
while(    ((-(J-Jold)>tol) && (iter>30))...
       || (iter>maxIter) )   % Safety brake
    % 1. Assign examples to the nearest centroid
    %------------------------------------------------------------
    for i = 1:m % Fix an example...
        temp = zeros(1,k);
        for j = 1:k % Compute distance to j-th centroid
            temp(j) = norm(examples(:,i)-us(:,j));
        end
        [~, minj] = min(temp);
        cs(i) = minj;
    end
    % 2. Update centroids
    %------------------------------------------------------------
    for j = 1:k % Fix a centroid
        sieve = logical(cs==j); % Subset of examples belonging
                                %   to j-th centroid
        us(:,j) = mean(examples(:,sieve),2);
    end
    % 3. Update loop conditions
    %------------------------------------------------------------
    iter = iter + 1;
    Jold = J;
    J = 0;
    for i = 1:m
        J = J + norm(examples(:,i)-us(:,cs(i)))^2;
    end
end

Published with MATLAB® 7.12