FIX multiply by random number < 0.5 for BorderlineSMOTE-2 #1027
Conversation
|
@solegalli Do the fix looks okay to you? I still find that the paper is ambiguous. Indeed, you could apply a random number for each feature of the Currently, we generate samples on the segments (the SMOTE paper is rather puzzling about this). Any thoughts? |
Co-authored-by: Soledad Galli <solegalli@protonmail.com>
Co-authored-by: Soledad Galli <solegalli@protonmail.com>
|
|
||
| Returns | ||
| ------- | ||
| X_new : {ndarray, sparse matrix} of shape (n_samples, n_features) | ||
| Synthetically generated samples. | ||
| """ | ||
| diffs = nn_data[nn_num[rows, cols]] - X[rows] | ||
| if y is not None: # only entering for BorderlineSMOTE-2 |
There was a problem hiding this comment.
LGTM, clever implementation.
There was a problem hiding this comment.
On second thoughts, would it not be enough to just half the diffs if we are multiplying it by steps in 186/188?
There was a problem hiding this comment.
The paper states to use a random number. If we take half, we always use 0.5.
|
Hey @glemaitre I agree the paper is vague for bordeline 2. The current code reflects what I also understand from the article. Thank you! |
Address issue raised in https://github.com/scikit-learn-contrib/imbalanced-learn/pull/1023/files#r1259422379
We additionally detect if the pair of samples used to generate samples are from different classes. In this case, we multiplied by a random number between 0 and 0.5.