Hey there! Let's work through your LDA implementation hurdles—since PCA gave you overlapping visualizations, LDA's supervised approach should help with better class separation, and I'll walk you through the correct steps for MATLAB (since you referenced the stats toolbox docs).

Key First Note: LDA Needs Labels! #

Unlike PCA (unsupervised), LDA is a supervised method—you must have category labels for your 100 samples. If you haven't been using labels yet, that's probably where you got stuck. Let's assume you have:

• X: Your (100,61) feature matrix
• Y: A 100x1 vector of category labels (can be categorical, numeric, or string)

Step 1: Prepare Your Data (Optional but Recommended) #

If your features have very different scales, standardize them first to avoid bias toward high-magnitude features:

X = zscore(X); % Centers features to mean 0, variance 1

Step 2: Train the LDA Classifier #

MATLAB has two common functions for LDA, depending on your version:

For newer MATLAB versions (R2016a+): Use fitcdiscr

% Train LDA model
ldaModel = fitcdiscr(X, Y);

% Extract the projected (reduced-dimension) scores
ldaScores = ldaModel.Scores;
% The number of LDA components is min(number of features, number of classes - 1)
% So if you have 2 classes, this will be 1D; 3 classes gives 2D (perfect for plotting)

For older MATLAB versions: Use lda

[coeff, ldaScores, latent] = lda(X, Y);
% ldaScores is your reduced-dimension data, with columns equal to (number of classes - 1)

Step 3: Visualize the LDA Results #

Now plot the reduced data to check for class separation:

If you have 2 classes (1D projection):

gscatter(ldaScores(:,1), zeros(size(ldaScores,1),1), Y);
title('LDA 1D Projection of Your Dataset');
xlabel('LDA Component 1');
grid on;

If you have 3+ classes (2D projection):

gscatter(ldaScores(:,1), ldaScores(:,2), Y);
title('LDA 2D Projection of Your Dataset');
xlabel('LDA Component 1');
ylabel('LDA Component 2');
grid on;

Common Pitfalls to Fix Your Issues #

• Forgetting category labels: If you tried using LDA without Y, MATLAB will throw an error—make sure you pass both your feature matrix and labels.
• Using fisheriris directly: That's just a sample dataset! You need to replace the sample data (meas and species) with your own X and Y.
• Too few classes: LDA requires at least 2 distinct classes to compute class separation—if all your samples are in one class, LDA won't work.

内容的提问来源于stack exchange，提问作者user9430368