Select Features — selectFeatures

Remove numeric fields using variance, correlation or relevance to a target.

Use cases

• Remove constant or low-variance numeric columns.
• Compare a smaller set of numeric predictors during exploration.

What it does

variance drops numeric columns that do not exceed the variance threshold. correlation computes absolute pairwise correlation and drops the later column of pairs above the threshold, preserving original ordering. Both examine all numeric columns and leave non-numeric columns in place.

kbest considers numeric columns other than the target, keeps up to k selected features, and retains the target and non-feature columns in their original order. Its scorer is f_regression for a numeric target with more than 20 distinct values; otherwise it is f_classif. The node refits on every input and emits a dataframe, not a fitted selector.

Configuration

Generated Python code

Download the customer sample CSV and save it as customers.csv beside the script. These are complete exports from File Input → this node → File Output; additional preparation steps are stated below.

The flow first uses Drop columns to remove customer_id. Then kbest selects two predictors of churn and writes select-features.csv. It keeps the 16 rows, the target and two numeric feature columns.

The node configuration is:

{
  "method": "kbest",
  "target_column": "churn",
  "k": 2
}

Both flow engines run this ML step through pandas/scikit-learn. The Polars script includes the actual conversion into pandas and back into Polars.

pandas

import pandas as pd
from sklearn.feature_selection import SelectKBest, f_classif, f_regression

df_customers = pd.read_csv('customers.csv')
df_customers = df_customers.drop(columns=['customer_id'])
_features = [c for c in df_customers.select_dtypes(include='number').columns if c != 'churn']
_scorer = f_regression if (pd.api.types.is_numeric_dtype(df_customers['churn']) and df_customers['churn'].nunique() > 20) else f_classif
_sel = SelectKBest(_scorer, k=min(2, len(_features))).fit(df_customers[_features], df_customers['churn'])
_kept = [c for c, kp in zip(_features, _sel.get_support()) if kp]
df_customers = df_customers.drop(columns=[c for c in _features if c not in _kept])
df_customers.to_csv('select-features.csv', index=False)

Polars

import polars as pl
import pandas as pd
from sklearn.feature_selection import SelectKBest, f_classif, f_regression

df_customers = pl.read_csv('customers.csv')
df_customers = df_customers.drop(['customer_id'])
_eager_1 = df_customers.to_pandas()
_features = [c for c in _eager_1.select_dtypes(include='number').columns if c != 'churn']
_scorer = f_regression if (pd.api.types.is_numeric_dtype(_eager_1['churn']) and _eager_1['churn'].nunique() > 20) else f_classif
_sel = SelectKBest(_scorer, k=min(2, len(_features))).fit(_eager_1[_features], _eager_1['churn'])
_kept = [c for c, kp in zip(_features, _sel.get_support()) if kp]
_eager_2 = _eager_1.drop(columns=[c for c in _features if c not in _kept])
df_1 = pl.from_pandas(_eager_2)
df_1.write_csv('select-features.csv')

Tips & common mistakes

• The catalog starts with variance threshold zero. Check the threshold when changing methods rather than assuming the editor applies the omitted-key correlation default.
• Variance/correlation do not distinguish numeric IDs or targets from features; either can be dropped.
• kbest chooses its scorer from target dtype and distinct-value count. A small continuous target can be treated as classification; inspect whether that is appropriate.
• Fit feature selection only on training rows when evaluating a model. This node does not persist a selector for a separate test branch; export and retain a fitted selector in your own pipeline when needed.
• If no numeric feature survives a variance threshold, scikit-learn raises an error; lower the threshold or change the input.

See also

• Train/Test Split · Reduce Dimensions
• Drop columns
• Feature-selection leakage guidance