id3 function D = ID3(train_features, train_targets, params, region) % Classify using Quinlan's ID3 algorithm % Inputs: % features - Train features % targets - Train targets % params - [Number of bins for the data, Percentage of incorrectly assigned samples at a node] % region - Decision region vector: [-x x -y y number_of_points] % % Outputs % D - Decision sufrace [Ni, M] = size(train_features); %Get parameters [Nbins, inc_node] = process_params(params); inc_node = inc_node*M/100; %For the decision region N = region(5); mx = ones(N,1) * linspace (region(1),region(2),N); my = linspace (region(3),region(4),N)' * ones(1,N); flatxy = [mx(:), my(:)]'; %Preprocessing [f, t, UW, m] = PCA(train_features, train_targets, Ni, region); train_features = UW * (train_features - m*ones(1,M));; flatxy = UW * (flatxy - m*ones(1,N^2));; %First, bin the data and the decision region data [H, binned_features]= high_histogram(train_features, Nbins, region); [H, binned_xy] = high_histogram(flatxy, Nbins, region); %Build the tree recursively disp('Building tree') tree = make_tree(binned_features, train_targets, inc_node, Nbins); %Make the decision region according to the tree disp('Building decision surface using the tree') targets = use_tree(binned_xy, 1:N^2, tree, Nbins, unique(train_targets)); D = reshape(targets,N,N); %END function targets = use_tree(features, indices, tree, Nbins, Uc) %Classify recursively using a tree targets = zeros(1, size(features,2)); if (size(features,1) == 1), %Only one dimension left, so work on it for i = 1:Nbins, in = indices(find(features(indices) == i)); if ~isempty(in), if isfinite(tree.child(i)), targets(in) = tree.child(i); else %No data was found in the training set for this bin, so choose it randomally n = 1 + floor(rand(1)*length(Uc)); targets(in) = Uc(n); end end end break end %This is not the last level of the tree, so: %First, find the dimension we are to work on dim = tree.split_dim; dims= find(~ismember(1:size(features,1), dim)); %And classify according to it for i = 1:Nbins, in = indices(find(features(dim, indices) == i)); targets = targets + use_tree(features(dims, :), in, tree.child(i), Nbins, Uc); end %END use_tree function tree = make_tree(features, targets, inc_node, Nbins) %Build a tree recursively [Ni, L] = size(features); Uc = unique(targets); %When to stop: If the dimension is one or the number of examples is small if ((Ni == 1) | (inc_node > L)), %Compute the children non-recursively for i = 1:Nbins, tree.split_dim = 0; indices = find(features == i); if ~isempty(indices), if (length(unique(targets(indices))) == 1), tree.child(i) = targets(indices(1)); else H = hist(targets(indices), Uc); [m, T] = max(H); tree.child(i) = Uc(T); end else tree.child(i) = inf; end end break end %Compute the node's I for i = 1:Ni, Pnode(i) = length(find(targets == Uc(i))) / L; end Inode = -sum(Pnode.*log(Pnode)/log(2)); %For each dimension, compute the gain ratio impurity delta_Ib = zeros(1, Ni); P = zeros(length(Uc), Nbins); for i = 1:Ni, for j = 1:length(Uc), for k = 1:Nbins, indices = find((targets == Uc(j)) & (features(i,:) == k)); P(j,k) = length(indices); end end Pk = sum(P); P = P/L; Pk = Pk/sum(Pk); info = sum(-P.*log(eps+P)/log(2)); delta_Ib(i) = (Inode-sum(Pk.*info))/-sum(Pk.*log(eps+Pk)/log(2)); end %Find the dimension minimizing delta_Ib [m, dim] = max(delta_Ib); %Split along the 'dim' dimension tree.split_dim = dim; dims = find(~ismember(1:Ni, dim)); for i = 1:Nbins, indices = find(features(dim, :) == i); tree.child(i) = make_tree(features(dims, indices), targets(indices), inc_node, Nbins); end