mljar-supervised
mljar-supervised copied to clipboard
Pass class cost matrix (weight) in AutoML fit
Please check if and how a cost matrix is supported in:
- scikit-learn https://stackoverflow.com/questions/25070910/does-sklearn-support-a-cost-matrix
- xgboost
- catboost
- lightgbm
- keras
This would be a huge improvement. I've been holding off with this library until this is ready.
Thanks!
@kmedved could you give an example of the cost matrix format? What is your use case?
I may be misunderstanding this issue (I came here from #154). I was thinking of adding a sample_weight
parameter, similar to what you see in scikit-learn, xgboost, catboost, lightgbm, etc... regression estimators. You would pass a column of sample weights along with your target variables, and features.
The use case is anytime you want to downweight some observations. This is common anytime you have different confidences in certain observations, such as if they're from a long time ago, or are just already rolled up in some way, so given rows represent different samples of data.