Modify Classifiers to Include Class Weighting?

[jentlapalm]

Hello sktime community!

Is there a way for me to modify or add a class_weight parameter to any of the sktime classifiers similar to the class_weight parameter included in many sklearn models?

I'm working on a classification project with imbalanced data (target variable is ~10% of data); the goal is to maximize recall of the target class.

The sktime classifiers I've tried have high accuracy, but achieve this by predicting nearly everything as part of the majority class.

[fkiraly]

Afaik the classification module does not support class weights currently.

You can try undersampling/oversampling though before you apply classifiers.

Perhaps some of the classification aficionados has a better suggestion.

[sssilvar]

I'm affraid I haven't seen this functionality in sktime yet. You could either resample using a workaround in imblearn. Example:

def _balance_classes(
    target: ArrayLike,
    ts: NDArray[np.float_],
) -> tuple[pd.DataFrame, NDArray[np.float_]]:
    """Balances class distribution by under-sampling majority classes."""
    rus = RandomUnderSampler()  # You could replace this by an Oversampler
    _, y_resampled = rus.fit_resample(metadata, target)
    ts_resampled = ts[y_resampled.index]
    metadata_resampled = metadata.loc[y_resampled.index]
    return metadata_resampled, ts_resampled

This will help you deal with the panel data sktime supports (talking about 3D arrays with shape (n_instances, n_variables, n_timepoints). It would be nice to have though a more advanced resampling like the iterative resamplers in torch or keras but it'd make the library too dependent on those as the workflow sktime seems to use is more like in sklearn where training data is immutable (Data panel -> pipeline).

[fkiraly]

hm, under/oversampling should not be too difficult to implement as a transformer component.

Integration with imblearn might be a nice project, to apply to classifiers, regressors, etc.

Would you like to open an issue with this suggestion, @sssilvar, so we can track it as a feature request?

[fkiraly]

also related to the "data augmentation" idea, FYI @benHeid, @astrogilda.

Under/Oversampling has the same signature-type as bootstrap augmenters.

[sssilvar]

Working on it !
Started bad though. I've extended it from BaseTransformer. I will adapt it.