File Exchange

image thumbnail

Kernel PCA and Pre-Image Reconstruction

version (6.9 MB) by Quan Wang
standard PCA, Gaussian kernel PCA, polynomial kernel PCA, pre-image reconstruction


Updated 02 Sep 2014

View License

In this package, we implement standard PCA, kernel PCA, and pre-image reconstruction of Gaussian kernel PCA. We also provide three demos: (1) Two concentric spheres embedding; (2) Face classification with PCA/kPCA; (3) Active shape models with kPCA.
Standard PCA is not optimized for very high dimensional data. But our kernel PCA implementation is very efficient, and has been used in many research projects.

Comments and Ratings (26)

Hello. I am trying to implement this code for my dataset. I am using Indian Pine dataset (AVIRIS 92AV3C) which is a 3 dimensional dataset size of 145x145x220. There are 145x145 pixels and 220 bandsor features and there are 16 classes. I have converted the dataset into 2 dimensional dataset. Code is given below.
load Indian_pines.mat;
[nr,nc,band]=size(indian_pines);% Size of the input data
totalpix=nr*nc; % total number of pixel
tddata=reshape(indian_pines,nr*nc,band); %converting from 3D to 2D data
DIST=distanceMatrix(tddata); %calculating distance matrix

when I am executing this line to calculate distance matrix it shows error something like this:
Error using repmat
Out of memory. Type HELP MEMORY for your options.

Error in distanceMatrix (line 15)
Can you please help me to sort out this problem?

Bran Sun

Hi, the pre-image is not looking correct when I try to run demo1. Please, try to run it at your end and check.

Hi Wang, I have a small doubt. How do we define our own kernel here? I want to use an rbf kernel.


kumud arora

Why the test data kernel is not centralized??


which type of data should be used to execute this code?


what is the form of the database to use


Hi, Wang!
For several reasons, I need to do the KNN classifier after reducing the dimension of training and sampling data. When the type is guassian, it seems that the eigenvector will be complex number so that KNN is invaild. How can I solve this problem? Thanks so much.


Yan wang


please help us to do fault detection in nonlinear systems with kernel pca. because i'm a starter in this field.

Quan Wang

Hi Cheung,

If you compute ||a-b||, you will have to use loops. MATLAB loops are slow, and distanceMatrix avoids using loops, thus is very fast.




Quan Wang

Dear all,

We have updated the code and the document. The code now generates exactly the same results as shown in the document. I am sorry for the confusion in previous versions of this package. The current version has lots of significant improvements.


When running demo file why the 3rd fig is in 2d but not in 3d as plot3() is used to plot variable in 3d

Yan Ou


So Good!


Quan Wang

Beside, I myself am using this code package (v2.0) for a number of research projects. I am pretty sure the code works well and has been well optimized.

Quan Wang

Hi Kris. The "paper" (actually a course project report) was using unordered eigenvalues, while in the updated code I have decreasingly ordered eigenvalues. So the code is more "correct" in a scientific sense. If you want to generate the same results as the report, you can uncomment
"% eigValue=eigValue(1:min(size(X)));"
in kPCA.m.

Hope this helps!

Kris Villez

Dear Quan Wang,

Thanks for sharing your code. However, I am not able to reproduce the results displayed in Figure 3 and 4 of your paper.

For Figure 3, it is not clear what the order of the applied polynomial is.

For Figure 4, the two features are highly correlated while the should in fact be uncorrelated. I wondered if this is a mistake in the paper or in the code.

Thanks for your comments,


Fixed a fatal bug in pre-image reconstruction.

addpath('../code') in demo2

We replaces all demos, and the data used for the demo. We also updated the document to provide better illustration and better experiments. Now the code generates exactly the same results as shown in the paper.

The efficiency is optimized.

MATLAB Release Compatibility
Created with R2012b
Compatible with any release
Platform Compatibility
Windows macOS Linux