Dear Farhad Bulbul,

I hope you are doing in Matlab.

here is the code how to run LDA and get the classification accrucacy result:

classLDA=classify(test_features,train_features,trainlabels,'linear');

CA=classperf(testlabels,classLDA);

Did you try used another feature selection algorithm ? When you have 4096 features before and only one after, that could be mean you features are strongly correlated.

Dear sir, Yeah. You are right. I have used PCA.

QDA fits the data better than LDA, but has more parameters

to estimate.

Estimate the covariance matrix Σk separately for each class k,

k = 1, 2, ...,K.

I Quadratic discriminant function:

δk (x) = −

1

2

log |Σk | − 1

2

(x − µk )

T Σ

−1

k

(x − µk ) + log πk .

I Classification rule:

Gˆ (x) = arg max

k

δk (x) .

Hi,

I think you have a bug in your LDA code.   

Cheers

Hi Paul, I don't think so. Please check my code.

function [mappedX, mapping] = lda(X, labels, no_dims)

%LDA Perform the LDA algorithm

%

% [mappedX, mapping] = lda(X, labels, no_dims)

%

% The function runs LDA on a set of datapoints X. The variable

% no_dims sets the number of dimensions of the feature points in the

% embedded feature space (no_dims >= 1, default = 2). The maximum number

% for no_dims is the number of classes in your data minus 1.

% The function returns the coordinates of the low-dimensional data in

% mappedX. Furthermore, it returns information on the mapping in mapping.

%

%

% This file is part of the Matlab Toolbox for Dimensionality Reduction v0.6b.

% The toolbox can be obtained from http://www.cs.unimaas.nl/l.vandermaaten

% You are free to use, change, or redistribute this code in any way you

% want for non-commercial purposes. However, it is appreciated if you

% maintain the name of the original author.

%

% (C) Laurens van der Maaten

% Maastricht University, 2008

if ~exist('no_dims', 'var') || isempty(no_dims)

no_dims = 2;

end

% Make sure data is zero mean

mapping.mean = mean(X, 1);

X = bsxfun(@minus, X, mapping.mean);

% Make sure labels are nice

[classes, bar, labels] = unique(labels);

nc = length(classes);

% Intialize Sw

Sw = zeros(size(X, 2), size(X, 2));

% Compute total covariance matrix

St = cov(X);

% Sum over classes

for i=1:nc

% Get all instances with class i

cur_X = X(labels == i,:);

% Update within-class scatter

C = cov(cur_X);

p = size(cur_X, 1) / length(labels);

Sw = Sw + (p * C);

end

% Compute between class scatter

Sb = St - Sw;

Sb(isnan(Sb)) = 0; Sw(isnan(Sw)) = 0;

Sb(isinf(Sb)) = 0; Sw(isinf(Sw)) = 0;

% Make sure not to embed in too high dimension

if nc

@Farhad Bulbul ·

Sorry i can't check your code.

As a sanity check, why not try some other code?

http://uk.mathworks.com/help/stats/discriminant-analysis.html?nocookie=true#bs2owre

http://uk.mathworks.com/matlabcentral/fileexchange/29673-lda--linear-discriminant-analysis

to see if you get same result.

your code works. How about making sure your input data x and y. Although you got one feature as result of LDA, you can figure it out whether good or not in classification. Check your output parameter, mappedX using plot(mappedX, 'o') for instance. 

Hi there is also a great post here you can follow by python implementation. Its explains stepwise so you try in future

https://sebastianraschka.com/Articles/2014_python_lda.html