This document describes the steps to calculate between-class variance and within-class variance in linear discriminant analysis (LDA). It explains that LDA aims to maximize between-class variance, which represents the separation between class means, and minimize within-class variance, which represents the difference between class means and samples. The key steps are: (1) Calculating the mean of each class and total mean, (2) Calculating between-class variance as the distance between each class mean and total mean, (3) Calculating within-class variance as the difference between each sample and its class mean.
This document describes the steps to calculate between-class variance and within-class variance in linear discriminant analysis (LDA). It explains that LDA aims to maximize between-class variance, which represents the separation between class means, and minimize within-class variance, which represents the difference between class means and samples. The key steps are: (1) Calculating the mean of each class and total mean, (2) Calculating between-class variance as the distance between each class mean and total mean, (3) Calculating within-class variance as the difference between each sample and its class mean.
This document describes the steps to calculate between-class variance and within-class variance in linear discriminant analysis (LDA). It explains that LDA aims to maximize between-class variance, which represents the separation between class means, and minimize within-class variance, which represents the difference between class means and samples. The key steps are: (1) Calculating the mean of each class and total mean, (2) Calculating between-class variance as the distance between each class mean and total mean, (3) Calculating within-class variance as the difference between each sample and its class mean.
class variance. This section will explain these three 1 N c
ni µ= ∑ xi = ∑ µi (3) steps in detail, and then the full description of the LDA N i=1 i=1 N algorithm will be given. Figures (1 and 2) are used to visualize the steps of the LDA technique. where c represents the total number of classes (in our example c = 3). 2.2. Calculating the Between-Class Variance (SB ) The term (µi − µ)(µi − µ)T in Equation (1) repre- The between-class variance of the ith class (SBi ) rep- sents the separation distance between the mean of the resents the distance between the mean of the ith class ith class (µi ) and the total mean (µ), or simply it repre- (µi ) and the total mean (µ). LDA technique searches sents the between-class variance of the ith class (SBi ). for a lower-dimensional space, which is used to max- Substitute SBi into Equation (1) as follows: imize the between-class variance, or simply maxi- mize the separation distance between classes. To ex- (mi − m)2 = W T SBi W (4) plain how the between-class variance or the between- The total between-class variance is calculated as fol- class matrix (SB ) can be calculated, the following as- lows, (SB = ∑ci=1 ni SBi ). Figure (1, step (D)) shows first sumptions are made. Given the original data matrix how the between-class matrix of the first class (SB1 ) is X = {x1 , x2 , . . . , xN }, where xi represents the ith sam- calculated and then how the total between-class matrix ple, pattern, or observation and N is the total num- (SB ) is then calculated by adding all the between-class ber of samples. Each sample is represented by M fea- matrices of all classes. tures (xi ∈ R M ). In other words, each sample is repre- sented as a point in M-dimensional space. Assume the 2.3. Calculating the Within-Class Variance (SW ) data matrix is partitioned into c = 3 classes as follows, X = [ω1 , ω2 , ω3 ] as shown in Fig. (1, step (A)). Each The within-class variance of the ith class (SWi ) rep- class has five samples (i.e. n1 = n2 = n3 = 5), where ni resents the difference between the mean and the sam- represents the number of samples of the ith class. The ples of that class. LDA technique searches for a lower- total number of samples (N) is calculated as follows, dimensional space, which is used to minimize the dif- N = ∑3i=1 ni . ference between the projected mean (mi ) and the pro- To calculate the between-class variance (SB ), the jected samples of each class (W T xi ), or simply min- separation distance between different classes which is imizes the within-class variance [83,36]. The within- denoted by (mi − m) will be calculated as follows: class variance of each class (SW j ) is calculated as in Equation (5).
(mi − m)2 = (W T µi −W T µ)2 = W T (µi − µ)(µi − µ)T W
(1) ∑ (W T xi − m j )2 xi ∈ω j , j=1,...,c where mi represents the projection of the mean of the = ∑ (W T xi j −W T µ j )2 ith class and it is calculated as follows, mi = W T µi , xi ∈ω j , j=1,...,c where m is the projection of the total mean of all classes and it is calculated as follows, m = W T µ, W = ∑ W T (xi j − µ j )2W represents the transformation matrix of LDA1 , µi (1 × xi ∈ω j , j=1,...,c M) represents the mean of the ith class and it is com- = ∑ W T (xi j − µ j )(xi j − µ j )T W puted as in Equation (2), and µ(1×M) is the total mean xi ∈ω j , j=1,...,c of all classes and it can be computed as in Equation (3) [83,36]. Figure (1) shows the mean of each class and = ∑ W T SW j W the total mean in step (B and C), respectively. xi ∈ω j , j=1,...,c (5) 1 µj = xi (2) n j xi∑ ∈ω j From Equation (5), the within-class variance for each class can be calculated as follows, SW j = d Tj ∗d j = nj ∑i=1 (xi j − µ j )(xi j − µ j )T , where xi j represents the ith 1 The transformation matrix (W ) will be explained in Sect. 2.4 sample in the jth class as shown in Fig. (1, step (E,