Sklearn operators end the run on two configurations they could refuse or handle

[kz930]

What happened?

Both cases end the execution with an error from inside scikit-learn or from code generation, naming neither the column nor the field the user has to change.

A text column the user did not mean as a feature. The operators take every column except the target as features, fixed at X = table.drop(target, axis=1) (SklearnClassifierOpDesc.scala:43, SklearnTrainingOpDesc.scala:43, SklearnTestingOpDesc.scala:71). A dataset that carries a text column beside its numeric ones ends the run with ValueError: could not convert string to float. A timestamp column fails the same way with DTypePromotionError. Boolean and integer columns are fine. Nothing in the configuration narrows the feature set, so the only way through is a Projection upstream, which the error does not suggest. Turning on Count Vectorizer is not a workaround: it replaces the feature set with that one text column and discards the numeric ones.

Count Vectorizer on with no Text Attribute. text is not required, so the switch can be turned on with the field left empty. It then reaches code generation as a null and the operator's code becomes #EXCEPTION DURING CODE GENERATION, per PythonOperatorDescriptor.scala:36-44. text also accepts a column of any type, though CountVectorizer tokenizes documents and a numeric column raises from inside it.

Expected: a configuration that cannot work is refused while it is being written, and a column an estimator cannot fit is dropped and named rather than ending the run. text is a string column, and is required exactly when Count Vectorizer is on. Both are statable in the descriptor's schema, the second as the conditional required the Aggregate operator already uses. For the feature set, keeping the columns an estimator can fit and printing the ones left out follows what the rest of the codebase does with data it cannot use, where twenty-four visualization operators drop missing values before plotting.

How to reproduce?

Upload a CSV whose numeric columns sit beside a text column:

x1,x2,note,y
0.0,0.0,good great,0
0.1,0.2,great nice,0
1.0,1.0,bad awful,1
0.9,0.8,awful poor,1

Build CSV File Scan to Decision Tree from the Sklearn group, wire it to both the training and testing ports, set Target Attribute to y, leave Count Vectorizer off, and run. The execution stops on the ValueError. For the second case, turn Count Vectorizer on and leave Text Attribute empty.

Version/Branch

1.3.0-incubating-SNAPSHOT (main)

Relevant log output

ValueError: could not convert string to float: 'good great'