Linear Discriminant Analysis(LDA) - C-Classes
■ Linear Discriminant Analysis(LDA) - C-Classes
이전 글을 두개의 클래스를 판별하는 LDA에 대해서 알아 봤다. 그럼 여러개(C개)의 클래스를 어떻게 판별할 수 있을까? 접근은 2개의 클래스 판별 LDA 방법과 유사하다.
n-feature vectors를 가졌다면 다음과 같이 표현할 수 있다.
여기서 Y는 출력벡터, X는 입력벡터, W는 변환행렬이다.
- 즉 mxn입력 백터에 C개의 클래스를 LDA 분석을 하면 출력벡터는 c-1 by n개의 배열이 된다. 이에 중요한점은 각 클래스 마다 최적의 변환행렬을 계산해야 한다.
C개의 클래스를 가지는 입력 벡터를 LDA 분석하기 위한 단계는 다음과 같다.
1. 원래 데이터 차원에서 통계 계산
step 1: 클래스 내 분산 구하기
step 2: 클래스 간 분산 구하기
2. 투영된 데이터 차원에서 통계 계산
step 1: 평균 벡터 구하기
step 2: 클래스 내 분산 구하기
step 3: 클래스 간 분산 구하기
3. 목적행렬을 통한 최적 변환 행렬 찾기
는 최적 변환 행렬임
- 최적 변환 행렬은 일반적인 고유값 문제 해결로 얻을 수 있는 최고 고유값에 해당하는 고유벡터가 됨
- C개의 클래스는 C-1개의 변환행렬을 가짐
4. 차원 축소
추가적으로 LDA 접근은 두 가지 방법으로 나뉘어짐
- 클래스 종속: 각 클래스 마다 변환행렬 생성
- 클래스 독립: 하나의 변환행렬 생성
5. 예제
LDA에 쓸 3개의 클래스 샘플 생성
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 | %clear clear %dataset Generation %let the center of all classes be Mu = [5;5]; %%for the first class Mu1=[Mu(1)-3; Mu(2)+7]; CovM1 = [5 -1; -3 3]; %%for the second class Mu2=[Mu(1)-2.5; Mu(2)-3.5]; CovM2 = [4 0; 0 4]; %%for the third class Mu3=[Mu(1)+7; Mu(2)+5]; CovM3 = [3.5 1; 3 2.5]; %generating feature vectors using Box-Muller approach %Generate a random variable following uniform(0,1) having two features and %1000 feature vectors U=rand(2,1000); %Extracting from the generated uniform random variable two independent %uniform random variables; u1 = U(:,1:2:end); u2 = U(:,2:2:end); %Using u1 and u2, we will use Box-Muller method to generate the feature %vectors to follow standard normal X=sqrt((-2).*log(u1)).*(cos(2*pi.*u2)); clear u1 u2 U; %now ... Manipulating the generated features N(0,1) to following certain %mean and covariance orher than the standard normal %First we will change its variance then we will change its mean %Getting the eigen vectors and values of the covariance matrix [V,D] = eig(CovM1); % D is the eigen values matrix and V is the eigen vectors matrix newX=X; for j=1:size(X,2) newX(:,j)=V*sqrt(D)*X(:,j); end %changing its mean newX=newX+repmat(Mu1, 1, size(newX,2)); %now our dataset for the first class matrix will be X1 = newX; %each column is a feature vector, each row is a single feature %...do the same for the other two classes with difference means and covariance matrices [V,D] = eig(CovM2); newX=X; for j=1:size(X,2) newX(:,j)=V*sqrt(D)*X(:,j); end newX=newX+repmat(Mu2, 1, size(newX,2)); X2 = newX; [V,D] = eig(CovM3); newX=X; for j=1:size(X,2) newX(:,j)=V*sqrt(D)*X(:,j); end newX=newX+repmat(Mu3, 1, size(newX,2)); X3 = newX; %draw 2d scatter plot figure; hold on; plot(X1(1,:), X1(2,:), 'ro', 'markersize',10, 'linewidth', 3); plot(X2(1,:), X2(2,:), 'go', 'markersize',10, 'linewidth', 3); plot(X3(1,:), X3(2,:), 'bo', 'markersize',10, 'linewidth', 3); |
위 코드 수행시 아래와 같이 출력됨
LDA matlab 코드는 다음과 같다.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 | %% computing the LDA % class means Mu1= mean(X1')'; Mu3= mean(X2')'; Mu2= mean(X3')'; %overall mean Mu = (Mu1+Mu2+Mu3)./3; %class covariance matrices S1=cov(X1'); S2=cov(X2'); S3=cov(X3'); %within-class scatter matrix Sw=S1+S2+S3; %number of samples of each class N1=size(X1, 2); N2=size(X2, 2); N3=size(X3, 2); %between-class scatter matrix SB1=N1.*(Mu1-Mu)*(Mu1-Mu)'; SB2=N2.*(Mu2-Mu)*(Mu2-Mu)'; SB3=N3.*(Mu3-Mu)*(Mu3-Mu)'; SB=SB1+SB2+SB3; %computing the LDA projection invSw=inv(Sw); invSw_by_SB = invSw*SB; %getting the projection vectors %[V,D]=EIG(X) produces a diagonal matrix D of eigenvalues and a %full matrix V whose columns are the corresponding eigenvectors [V,D]=eig(invSw_by_SB); %the projcetion vectors - we will have at most C-1 projection vectors, %from which we can choose the most important ones ranked by their %corresponding eigen values ... lets investigate the two projection vectors W1=V(:,1); W2=V(:,2); %%lets visualize them... % we will plot the scatter plot to better visualize the features hfig=figure; axes1=axes('Parent',hfig,'FontWeight','bold','FontSize',12); hold('all'); %Create xLabel xlabel('X_1 - the first feature', 'FontWeight', 'bold', 'FontSize',12,'FontName', 'Garamond'); %Create yLabel ylabel('X_2 - the second feature', 'FontWeight', 'bold', 'FontSize',12,'FontName', 'Garamond'); %the fist class scatter(X1(1,:), X1(2,:),'r','LineWidth',2,'Parent',axes1); hold on %the second class scatter(X2(1,:), X2(2,:),'g','LineWidth',2,'Parent',axes1); hold on %the third class scatter(X3(1,:), X3(2,:),'b','LineWidth',2,'Parent',axes1); hold on %drawing the projection vectors %the first vector t=-10:25; line_x1 = t.*W1(1); line_y1 = t.*W1(1); %the second vector t=-5:20; line_x2 = t.*W2(1); line_y2 = t.*W2(2); plot(line_x1, line_y1, 'k-', 'LineWidth',3); hold on plot(line_x2, line_y2, 'm-', 'LineWidth',3); hold on %projection data samples along the projections axes %the first projection vector y1_w1 = W1'*X1; y2_w1 = W1'*X2; y3_w1 = W1'*X3; %projection limits minY=min([min(y1_w1), min(y2_w1), min(y3_w1)]); maxY=max([max(y1_w1), max(y2_w1), max(y3_w1)]); y_w1=minY:0.05:maxY; %for visualization lets compute the probability %density function of the classes after projection %the first class y1_w1_Mu = mean(y1_w1); y1_w1_sigma = std(y1_w1); y1_w1_pdf = mvnpdf(y1_w1',y1_w1_Mu,y1_w1_sigma); %the second class y2_w1_Mu = mean(y2_w1); y2_w1_sigma = std(y2_w1); y2_w1_pdf = mvnpdf(y1_w1',y2_w1_Mu,y2_w1_sigma); %the third class y3_w1_Mu = mean(y3_w1); y3_w1_sigma = std(y3_w1); y3_w1_pdf = mvnpdf(y1_w1',y3_w1_Mu,y3_w1_sigma); |
검은색이 고유값이 큰 고유벡터 값으로 판별되는 LD1 축이 되고, 다음 고유값에 따른 LD2 축이 보라색 선이 된다. 코드안에 차원축소를 한 데이터에 대해 PDF 분석 코드가 있다. 화면에 찍어야 하는데 그건 잘 모르겠다.
우선 LDA에 대해서 어떻게 접근해야 하는지 이제 좀 감이 잡힌다. 다시 LDA 전체 matlab 소스 코드를 첨부한다.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 | %clear clear %dataset Generation %let the center of all classes be Mu = [5;5]; %%for the first class Mu1=[Mu(1)-3; Mu(2)+7]; CovM1 = [5 -1; -3 3]; %%for the second class Mu2=[Mu(1)-2.5; Mu(2)-3.5]; CovM2 = [4 0; 0 4]; %%for the third class Mu3=[Mu(1)+7; Mu(2)+5]; CovM3 = [3.5 1; 3 2.5]; %generating feature vectors using Box-Muller approach %Generate a random variable following uniform(0,1) having two features and %1000 feature vectors U=rand(2,1000); %Extracting from the generated uniform random variable two independent %uniform random variables; u1 = U(:,1:2:end); u2 = U(:,2:2:end); %Using u1 and u2, we will use Box-Muller method to generate the feature %vectors to follow standard normal X=sqrt((-2).*log(u1)).*(cos(2*pi.*u2)); clear u1 u2 U; %now ... Manipulating the generated features N(0,1) to following certain %mean and covariance orher than the standard normal %First we will change its variance then we will change its mean %Getting the eigen vectors and values of the covariance matrix [V,D] = eig(CovM1); % D is the eigen values matrix and V is the eigen vectors matrix newX=X; for j=1:size(X,2) newX(:,j)=V*sqrt(D)*X(:,j); end %changing its mean newX=newX+repmat(Mu1, 1, size(newX,2)); %now our dataset for the first class matrix will be X1 = newX; %each column is a feature vector, each row is a single feature %...do the same for the other two classes with difference means and covariance matrices [V,D] = eig(CovM2); newX=X; for j=1:size(X,2) newX(:,j)=V*sqrt(D)*X(:,j); end newX=newX+repmat(Mu2, 1, size(newX,2)); X2 = newX; [V,D] = eig(CovM3); newX=X; for j=1:size(X,2) newX(:,j)=V*sqrt(D)*X(:,j); end newX=newX+repmat(Mu3, 1, size(newX,2)); X3 = newX; %draw 2d scatter plot figure; hold on; plot(X1(1,:), X1(2,:), 'ro', 'markersize',10, 'linewidth', 3); plot(X2(1,:), X2(2,:), 'go', 'markersize',10, 'linewidth', 3); plot(X3(1,:), X3(2,:), 'bo', 'markersize',10, 'linewidth', 3); %% computing the LDA % class means Mu1= mean(X1')'; Mu3= mean(X2')'; Mu2= mean(X3')'; %overall mean Mu = (Mu1+Mu2+Mu3)./3; %class covariance matrices S1=cov(X1'); S2=cov(X2'); S3=cov(X3'); %within-class scatter matrix Sw=S1+S2+S3; %number of samples of each class N1=size(X1, 2); N2=size(X2, 2); N3=size(X3, 2); %between-class scatter matrix SB1=N1.*(Mu1-Mu)*(Mu1-Mu)'; SB2=N2.*(Mu2-Mu)*(Mu2-Mu)'; SB3=N3.*(Mu3-Mu)*(Mu3-Mu)'; SB=SB1+SB2+SB3; %computing the LDA projection invSw=inv(Sw); invSw_by_SB = invSw*SB; %getting the projection vectors %[V,D]=EIG(X) produces a diagonal matrix D of eigenvalues and a %full matrix V whose columns are the corresponding eigenvectors [V,D]=eig(invSw_by_SB); %the projcetion vectors - we will have at most C-1 projection vectors, %from which we can choose the most important ones ranked by their %corresponding eigen values ... lets investigate the two projection vectors W1=V(:,1); W2=V(:,2); %%lets visualize them... % we will plot the scatter plot to better visualize the features hfig=figure; axes1=axes('Parent',hfig,'FontWeight','bold','FontSize',12); hold('all'); %Create xLabel xlabel('X_1 - the first feature', 'FontWeight', 'bold', 'FontSize',12,'FontName', 'Garamond'); %Create yLabel ylabel('X_2 - the second feature', 'FontWeight', 'bold', 'FontSize',12,'FontName', 'Garamond'); %the fist class scatter(X1(1,:), X1(2,:),'r','LineWidth',2,'Parent',axes1); hold on %the second class scatter(X2(1,:), X2(2,:),'g','LineWidth',2,'Parent',axes1); hold on %the third class scatter(X3(1,:), X3(2,:),'b','LineWidth',2,'Parent',axes1); hold on %drawing the projection vectors %the first vector t=-10:25; line_x1 = t.*W1(1); line_y1 = t.*W1(1); %the second vector t=-5:20; line_x2 = t.*W2(1); line_y2 = t.*W2(2); plot(line_x1, line_y1, 'k-', 'LineWidth',3); hold on plot(line_x2, line_y2, 'm-', 'LineWidth',3); hold on %projection data samples along the projections axes %the first projection vector y1_w1 = W1'*X1; y2_w1 = W1'*X2; y3_w1 = W1'*X3; %projection limits minY=min([min(y1_w1), min(y2_w1), min(y3_w1)]); maxY=max([max(y1_w1), max(y2_w1), max(y3_w1)]); y_w1=minY:0.05:maxY; %for visualization lets compute the probability %density function of the classes after projection %the first class y1_w1_Mu = mean(y1_w1); y1_w1_sigma = std(y1_w1); y1_w1_pdf = mvnpdf(y1_w1',y1_w1_Mu,y1_w1_sigma); %the second class y2_w1_Mu = mean(y2_w1); y2_w1_sigma = std(y2_w1); y2_w1_pdf = mvnpdf(y1_w1',y2_w1_Mu,y2_w1_sigma); %the third class y3_w1_Mu = mean(y3_w1); y3_w1_sigma = std(y3_w1); y3_w1_pdf = mvnpdf(y1_w1',y3_w1_Mu,y3_w1_sigma); |
[1] http://www.di.univr.it/documenti/OccorrenzaIns/matdid/matdid437773.pdf
[2] http://www.bytefish.de/blog/pca_lda_with_gnu_octave/