Idea:
Create predictor class TargetMeanPrediction or similar name with methods fit and predict.
- Fit - learns transformation
- predict - returns the mean target value per observation
Output:
- This transformer will automatically output the mean value of the target per category if variable is categorical (we have an encoder for this).
- If variable is numerical, it will first discretize it (we have discretizers for this, equal width and frequency, user selects) and then replace by the target mean.
The reason to create a predictor class is that then, we can use it with cross_validate and cross_val_score, in the main selector function.
Things to consider:
Then we need to re-code the class SelectByTargetMeanPerformance to call our predictor, and use it with cross-validate to return the important features. The advantage of using corss_validate is not just the cross_validation, which offers a less biased score, but it also allows the use of other metrics, not just roc and r2 as what we have at the moment.
Idea:
Create predictor class TargetMeanPrediction or similar name with methods fit and predict.
Output:
The reason to create a predictor class is that then, we can use it with cross_validate and cross_val_score, in the main selector function.
Things to consider:
Then we need to re-code the class SelectByTargetMeanPerformance to call our predictor, and use it with cross-validate to return the important features. The advantage of using corss_validate is not just the cross_validation, which offers a less biased score, but it also allows the use of other metrics, not just roc and r2 as what we have at the moment.