6 Comments
User's avatar
Giorgio Borelli's avatar

A limitation is a derivative of what you mentioned: how to pick a proper kernel to make the data linearly separable? How do we check we have made our data set linearly separable?

Expand full comment
Avi Chawla's avatar

Rightly said, Giorgio :)

I vividly remember looking around for the answers myself once I was stuck. There are some great answers here if you wish to read: https://stats.stackexchange.com/questions/131142/how-to-choose-a-kernel-for-kernel-pca.

Expand full comment
Giorgio Borelli's avatar

Thank you, good links

Expand full comment
Giorgio Borelli's avatar

You put in admirable efforts to offer a new angle on data science, each day! Well done

Expand full comment
Avi Chawla's avatar

Thanks for appreciating, Giorgio. Really appreciate the kind words :)

Expand full comment
Julius Nyambok's avatar

I beg to differ a bit.

PCA can sometimes make linear classification harder if the class separation information is not aligned with the principal components. This is not a "failure" of PCA, but rather a mismatch between PCA's objective (variance preservation) and the downstream task's requirement (class separation).

Expand full comment