Implementation of Logistic Regression using Matlab

MyLogisticRegression.m

% MyLogisticRegression
%
% by Faisal Wirakusuma
% Implementation of Logistic Regression, inspired by lecture materials from Andrew Ng
% http://openclassroom.stanford.edu/MainFolder/DocumentPage.php?course=MachineLearning&doc=exercises/ex4/ex4.html
% 

classdef MyLogisticRegression < handle 
    
    properties
        theta;
        sizem;
        sizen;
        %function theta(x), a sigmoid func
        g = inline('1.0 ./ (1.0 + exp(-z))');
        %default iterations parameter
        iterations = 5;
        %default learning rate parameter
        learningrate = 0.1;
        %convergence indicator J
        J;
        %error
        err = 0;
        %accuracy
        acc = 0;
        %predictedLabels;
        trainingtime;
    end
    
    methods
        
        function this = setlearningrate(lrate)
            this.learningrate = lrate;
        end
        
        function this = setiterations(iter)
            this.iterations = iter;
        end
        
        function this = train(this,data,labels)
            tic
            %convergence indicator J
            this.J = zeros(this.iterations, 1);
   
            [this.sizem, this.sizen] = size(data);
            temp_data = [ones(this.sizem,1), data];
            
            %find data and labels
            label_one = find(labels);
            label_zero = find(labels==0);
            
            %initialize theta
            this.theta = zeros(this.sizen+1, 1);
            
            for i = 1:this.iterations
                % Hypothesis function h(x)
                h = this.g(temp_data * this.theta);
                
                % Gradient
                grdnt = (1/this.sizem) .* temp_data' * (h-labels);
                % Hessian
                hessian = (1/this.sizem) .* temp_data' * diag(h) * diag(1-h) * temp_data;
                
                % Update J to prove convergence 
                this.J(i) = (1/this.sizem) * sum(-labels.*log(h) - (1-labels) .* log(1-h));
                
                % Update theta, calculate partial derivative of gradient
                
                % using Newton's method :
                %this.theta = this.theta - this.learningrate * (hessian\grdnt);
                
                gradForGD = 0.0;
                % using Gradient Descent :
                for j = 1:this.sizem
                    gradForGD = gradForGD + ( h(j) - labels(j) ) * temp_data(j, :)';
                end
                % gradient = nx1 column vector
                gradForGD = (1/this.sizem) * gradForGD;
                this.theta = this.theta - this.learningrate * gradForGD;
                
            end
            
            this.trainingtime =  toc;
        end%train
        
        function guess = test(this, test_data, test_labels)
            % Do the prediction from test data
            [test_m test_n] = size(test_data);
            test_data = [ones(test_m,1), test_data];
            % Predicts test data with sigmoid function
            preds = this.g(test_data*this.theta);
            
            % Assign predicted value to labels
            predictedLabels = zeros(length(preds),1);
            for i = 1:length(preds)
                if (preds(i)>0.5)
                    predictedLabels(i) = 1;
                else
                    predictedLabels(i) = 0;
                end
            end
            
            % Calculate error and accuracy
            % sum(y==y')/length(y)
            this.acc = sum(test_labels==predictedLabels)/length(test_labels);
            this.err = 1 - this.acc;
            
            %guess = results(predictedLabels);
            %if(exist('labels','var')
            %    guess.addtruelabels(labels);
            %end
            % Return the PREDICTIONS
            guess = predictedLabels;            
        end%test
        
    end%methods
    
end%classdef