20
$\begingroup$

Am i misunderstanding something. This is my code

using sklearn

import numpy as np
import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D
from sklearn import decomposition
from sklearn import datasets
from sklearn.preprocessing import StandardScaler

pca = decomposition.PCA(n_components=3)

x = np.array([
        [0.387,4878, 5.42],
        [0.723,12104,5.25],
        [1,12756,5.52],
        [1.524,6787,3.94],
    ])
pca.fit_transform(x)

Output:

array([[ -4.25324997e+03,  -8.41288672e-01,  -8.37858943e-03],
   [  2.97275001e+03,  -1.25977271e-01,   1.82476780e-01],
   [  3.62475003e+03,  -1.56843494e-01,  -1.65224286e-01],
   [ -2.34425007e+03,   1.12410944e+00,  -8.87390454e-03]])

Using numpy methods

x_std = StandardScaler().fit_transform(x)
cov = np.cov(x_std.T)
ev , eig = np.linalg.eig(cov)
a = eig.dot(x_std.T)

Output

array([[ 0.06406894,  0.94063993, -1.62373172],
   [-0.35357757,  0.7509653 ,  0.63365168],
   [ 0.29312477,  0.6710958 ,  1.11766206],
   [-0.00361615, -2.36270102, -0.12758202]])
I have kept all 3 components but it doesnt seem to allow me to retain my original data.

May i know why is it so?

If i want to obtain back my original matrix what should i do?

$\endgroup$
  • $\begingroup$ Your numpy code is wrong IMHO (also, it uses X which is not defined). Recheck your math. $\endgroup$ – Anony-Mousse Sep 20 '16 at 6:12
  • $\begingroup$ I am using ipython notebook so i could only copy by cells. My maths is wrong? Which part @Anony-Mousse $\endgroup$ – aceminer Sep 20 '16 at 8:41
  • $\begingroup$ @Anony-Mousse Yes i realised my error but it still doesnt match up $\endgroup$ – aceminer Sep 20 '16 at 8:50
  • $\begingroup$ @aceminer I am curious why you calculate covariance matrix of x_std.T, not x_std? $\endgroup$ – Evgeni Nabokov Mar 3 at 18:35
  • $\begingroup$ @EvgeniNabokov its been too long. Sry i cannot remember already $\endgroup$ – aceminer Mar 6 at 6:58
21
$\begingroup$

The difference is because decomposition.PCA does not standardize your variables before doing PCA, whereas in your manual computation you call StandardScaler to do the standardization. Hence, you are observing this difference: PCA on correlation or covariance?

If you replace

pca.fit_transform(x)

with

x_std = StandardScaler().fit_transform(x)
pca.fit_transform(x_std)

you will get the same result as with manual computation...

...but only up to the order of the PCs. That is because when you run

ev , eig = np.linalg.eig(cov)

you get eigenvalues not necessarily in the decreasing order. I get

array([ 0.07168571,  2.49382602,  1.43448827])

So you will want to order them manually. Sklearn does that for you.


Regarding reconstructing original variables, please see How to reverse PCA and reconstruct original variables from several principal components?

$\endgroup$
  • $\begingroup$ Just would like to check. Is it really necessary to standardize the matrix by its standard deviation? I seen examples where they do not do it $\endgroup$ – aceminer Sep 20 '16 at 13:50
  • $\begingroup$ It is not necessary, it is just one way of doing it. See the link that I put in the first paragraph: stats.stackexchange.com/questions/53 -- it really is all about this question. If you standardize, you do PCA on correlations. If you do not, you do PCA on covariances. $\endgroup$ – amoeba says Reinstate Monica Sep 20 '16 at 14:07
8
$\begingroup$

Here is a nice implementation with discussion and explanation of PCA in python. This implementation leads to the same result as the scikit PCA. This is another indicator that your PCA is wrong.

import numpy as np
from scipy import linalg as LA

x = np.array([
        [0.387,4878, 5.42],
        [0.723,12104,5.25],
        [1,12756,5.52],
        [1.524,6787,3.94],
    ])

#centering the data
x -= np.mean(x, axis = 0)  

cov = np.cov(x, rowvar = False)

evals , evecs = LA.eigh(cov)

you need to sort the eigenvalues (and eigenvectors accordingly) descending

idx = np.argsort(evals)[::-1]
evecs = evecs[:,idx]
evals = evals[idx]

a = np.dot(x, evecs) 

Generally, I recommend you to you check your code by implementation a simple example (as simple as possible) and calculating by hand the correct resuls (and intermediate results). This helps you to identify the problem.

$\endgroup$
  • 1
    $\begingroup$ Love this answer. It solved my problem! $\endgroup$ – Jinhua Wang Jul 12 '18 at 15:28

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.