1
$\begingroup$

Are there any reliable libraries or methods for tabular/structured data (with numerical and categorical features) augmentation? Could you share some?

Basically I believe inventing/augmenting more data could help to improve model performance.

$\endgroup$

2 Answers 2

1
$\begingroup$

It's hard in general, because you don't know the joint distribution of the data, let alone joint distribution with the label -- or else you wouldn't need a classifier. Without that you can't confidently sample new, valid instances.

In the image case it's 'easy' because we know certainly that a rotation or shear or scale of an image produces another valid image, and we know its label is the same.

If you're willing to make the assumption that points in your input space between two instances with the same label are also a) valid inputs and b) have the same label, then you can apply things like SMOTE (https://imbalanced-learn.readthedocs.io/en/stable/generated/imblearn.over_sampling.SMOTE.html) to generate more data according to that assumption. If that assumption is wrong, it'll hurt performance.

$\endgroup$
1
  • $\begingroup$ Thanks for the recommendation Sean! I guess it wouldn't hurt to test this out :) $\endgroup$
    – ValdemarT
    Commented Feb 11, 2020 at 12:28
0
$\begingroup$

There is a sophisticated way of data augmentation on tabular data that I found on a Kaggle winning solution:

  1. Add noise to the data by randomly replacing some features of each row with other rows.
  2. Use a DAE (Denoising Auto-Encoder) to rebuild the data.
$\endgroup$

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.