Add ability to pass sample weights for training

[pplonski]

There should be an option to pass weights for training. Weights should be used in metrics as well.

[diogosilva30]

Do you mean class_weights? So it pays more "attention" to a certain class in a classification task?

[pplonski]

I mean sample weights. For class weights, I've created separate issue #155

[pplonski]

@kmedved this can be added to the next major release. Would you like to help? Even testing or providing use cases would be nice.

[kmedved]

I would be happy to help with testing or use cases; actual code contribution is over my head unfortunately.

[pplonski]

User can pass sample_weight in the fit():

automl = AutoML()
automl.fit(X, y, sample_weight)

The default for sample_weight is None. All metrics (eval_metric and additional metrics) use sample_weight if available.

The algorithms that don't support sample_weight:

• MLP Classfier and Regressor (docs)
• kNN Classifier and Regressor (docs)

[pplonski]

@kmedved I just finished sample_weight functionality. It is available in the dev branch. If you have use cases for testing, I'll be happy to hear feedback from you!

[kmedved]

Thanks @pplonski - I will do some testing. Thanks!

[alitirmizi23]

@pplonski does adding sample_weight in .fit() cause training to pay more attention to heavily weighted samples or is it just for calculating metrics?

[pplonski]

@alitirmizi23 you are right, the higher the weight the more attention during the training.