5
$\begingroup$

I have a large data-set with >100 features. Can I use PCA to reduce the dimension of each group of features instead of t applying PCA to the whole data-set?

For example, if I have 10 demographic features and 10 work-place related features in my dataset, can I apply PCA to the demographic features, extract 2 PCs, do the same to the workplace features and do the final analysis on 4 extracted PCs? How does my described method compare with applying PCA to the entire dataset at once statistically? I only intend to use this approach for an slight increase in interpretability. Is it worth it? what are the pros and cons?

The features in each feature groups are more or less correlated but not necessarily.

Improve this question
$\endgroup$

1 Answer 1

Reset to default
9
$\begingroup$

You can certainly do this. For pros, as you have mentioned, the new features may be more interpretable/explainable. With this approach, you'll possibly end up selecting more features to describe the same amount of variance as you are limiting certain linear combinations of cross group features, though this is not necessarily a bad thing.

Improve this answer
$\endgroup$
2
  • 4
    $\begingroup$ +1 I have done exactly this when working with large numbers of chemical analytes. They tend to fall into groups of common chemical, biological, and/or physical behaviors, such as metals, chlorinated solvents, polynuclear aromatics, and so on; and often each group benefits from its own separate analysis. $\endgroup$
    – whuber ♦
    Commented Feb 28, 2022 at 19:19
  • 1
    $\begingroup$ @whuber I similarly was able to separate some climate data into snow-related principal components and rain-related principal components. $\endgroup$
    – Galen
    Commented Feb 28, 2022 at 23:00

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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