What is right way to create a feature when there are missing values

[gengyabc]

I want to construct a feature that should be rounded like below, but I "cannot convert float NaN to integer".

but if I use "Preprocess", I cannot construct my feature

Can I use "Test and Score" in this situation? Or what should I do to prevent testing set leaking to training set?

BTW, is below one of right way?

[janezd]

The error message looks strange ... but can your MonthlyCharges equal 0? Does round(TotalCharges/MonthlyCharges) if MonthlyCharges > 0 else 0 work better?

Note that you probably don't want to use 0, as above. Perhaps you should better use Select Rows to remove any rows with zero monthly charges.

[ajdapretnar]

That's not the case. I've tried it on heart-disease and the two error messages are different. ZeroDivisionError for dividing by 0 and ValueError when missing values are in the data. I wasn't able to write a normal-looking statement that would handle both.

[janezd]

,,, if monthlyCharges == monthlyCharges != 0 ... is False when monthlyCharges is either nan or 0. :)

[gengyabc]

There is no 0 in MonthlyCharges.
I want to impute the missing values before create the new feature but I don't find the proper way to do it if I want to use "Test and Score".

The similar situation is what if I want to use tree based imputer? "Preprocess" widget cannot do it. The only way is using "Data sampler" with "Prediction"?

In these situations, is there any way to use "Test and Score" ?

[janezd]

You can connect Preprocess directly to Test and Score. Preprocessors will be "trained" on training data only, and the used on testing data.

For instance, if you impute mean values, the mean will be computed from training data only, but imputation will also be done on testing data. This is Ok, because no information leaks from testing data to the model.

[gengyabc]

But how can I impute before "feature constructor"?

[janezd]

Ah, of course, I see.

Can you try using round(TotalCharges/MonthlyCharges) if MonthlyCharges == MonthlyCharges > 0 else 0 in Feature Constructor? This should handle missing values, I guess.

[gengyabc]

yes, this solution kind of solved the problem, though not the ideal way. Thank you very much.