Feature Importance — `featureImportance`

Inspect which features a trained tree or linear model uses most.

What it does

Connect a Train node's purple model output to this node's only input, model. It loads the trained artifact and emits an out frame containing feature_name, importance and rank, sorted by decreasing importance. Ranks start at 1. It does not accept a data wire or train another model. An unfitted Classifier/Regressor Model definition has no usable model_uri and cannot provide importances.

Tree estimators supply feature_importances_. Linear estimators supply absolute coef_ values; multiclass coefficients are averaged in absolute value across classes. Models with neither attribute, such as KNN or an RBF SVM, raise an error. When preprocessing expands a category, the result ranks the transformed features separately rather than combining them into one score.

Names come from fitted preprocessing or the model's feature names. If names cannot be recovered, runtime uses generic names such as feature_0. Preview passes the model-reference frame through without loading the model or extracting importance; run the flow to obtain the ranked table.

Configuration

Config key	Type	Required	Description
`top_n`	positive integer or null	No	Default `null`; return every feature, or the first N ranked rows

A limit greater than the feature count returns every available feature. Tree importances and absolute linear coefficients have different scales; coefficients are not automatically normalized. Importance does not establish causality, and correlated features can share or obscure each other's scores.

Generated Python code

Download the customer sample CSV and save it as customers.csv beside the script. Run with Ciaren installed and its ML dependencies available. Both engines use pandas/scikit-learn at the ML boundary; the Polars export converts back for the CSV output.

The sample has 16 rows. Train/Test Split uses test_size=0.25 and seed=42, so training sees 12 rows and four distinct rows remain held out. The scripts fit preprocessing on training rows only. In the editor, Train logs an MLflow model artifact and consumers load its reference. These exports fit a pipeline in memory and do not save a model or create an MLflow run.

This example trains a 20-tree random forest on churn, using tenure, monthly_charges and support_calls as features. Train/Test Split is stratified by churn; only the 12 training rows fit the model. The four held-out rows remain unused in importance extraction. top_n=2 writes two ranked rows to feature-importance.csv.

The consumer configuration is:

json

{
  "top_n": 2
}

The Train configuration is:

json

{
  "model_type": "random_forest_classifier",
  "target_column": "churn",
  "feature_columns": [
    "tenure",
    "monthly_charges",
    "support_calls"
  ],
  "hyperparameters": {
    "n_estimators": 20
  },
  "seed": 42
}

pandas

python

import pandas as pd
import joblib
from sklearn.compose import ColumnTransformer
from sklearn.ensemble import RandomForestClassifier
from sklearn.impute import SimpleImputer
from sklearn.model_selection import train_test_split
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import OneHotEncoder
from sklearn.preprocessing import StandardScaler

df_customers = pd.read_csv('customers.csv')
df_1, df_2 = train_test_split(df_customers, test_size=0.25, random_state=42, stratify=df_customers['churn'])
df_1 = df_1.reset_index(drop=True)
df_2 = df_2.reset_index(drop=True)
_features = ['tenure', 'monthly_charges', 'support_calls']
_X = df_1[_features]
_numeric = [c for c in _features if pd.api.types.is_numeric_dtype(_X[c])]
_categorical = [c for c in _features if c not in _numeric]
_transformers = []
if _numeric:
    _transformers.append(('num', Pipeline([('impute', SimpleImputer(strategy='median')), ('scale', StandardScaler())]), _numeric))
if _categorical:
    _transformers.append(('cat', Pipeline([('impute', SimpleImputer(strategy='most_frequent')), ('encode', OneHotEncoder(handle_unknown='ignore'))]), _categorical))
_preprocessor = ColumnTransformer(_transformers, remainder='drop')
_y = df_1['churn']
df_3 = Pipeline([('preprocessor', _preprocessor), ('model', RandomForestClassifier(n_estimators=20, random_state=42))])
df_3.fit(_X, _y)
_est = df_3.named_steps.get('model', df_3) if hasattr(df_3, 'named_steps') else df_3
_imp = getattr(_est, 'feature_importances_', None)
if _imp is None:
    _imp = abs(_est.coef_).mean(axis=0) if _est.coef_.ndim > 1 else abs(_est.coef_)
if hasattr(df_3, 'named_steps') and 'preprocessor' in df_3.named_steps:
    _raw = [str(c) for c in df_3[:-1].get_feature_names_out()]
    _names = [next((c[len(p):] for p in ('num__', 'cat__', 'remainder__') if c.startswith(p)), c) for c in _raw]
else:
    _fn = getattr(_est, 'feature_names_in_', None)
    _names = [str(c) for c in _fn] if _fn is not None else [f'feature_{i}' for i in range(len(_imp))]
df_3 = pd.DataFrame({'feature_name': _names, 'importance': _imp}).sort_values('importance', ascending=False).reset_index(drop=True)
df_3['rank'] = df_3.index + 1
df_3 = df_3.head(2)
df_3.to_csv('feature-importance.csv', index=False)

polars

python

import polars as pl
import joblib
import pandas as pd
from sklearn.compose import ColumnTransformer
from sklearn.ensemble import RandomForestClassifier
from sklearn.impute import SimpleImputer
from sklearn.model_selection import train_test_split
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import OneHotEncoder
from sklearn.preprocessing import StandardScaler

df_customers = pl.read_csv('customers.csv')
_eager_1 = df_customers.to_pandas()
_eager_2, _eager_3 = train_test_split(_eager_1, test_size=0.25, random_state=42, stratify=_eager_1['churn'])
_eager_2 = _eager_2.reset_index(drop=True)
_eager_3 = _eager_3.reset_index(drop=True)
df_1 = pl.from_pandas(_eager_2)
df_2 = pl.from_pandas(_eager_3)
_eager_4 = df_1.to_pandas()
_features = ['tenure', 'monthly_charges', 'support_calls']
_X = _eager_4[_features]
_numeric = [c for c in _features if pd.api.types.is_numeric_dtype(_X[c])]
_categorical = [c for c in _features if c not in _numeric]
_transformers = []
if _numeric:
    _transformers.append(('num', Pipeline([('impute', SimpleImputer(strategy='median')), ('scale', StandardScaler())]), _numeric))
if _categorical:
    _transformers.append(('cat', Pipeline([('impute', SimpleImputer(strategy='most_frequent')), ('encode', OneHotEncoder(handle_unknown='ignore'))]), _categorical))
_preprocessor = ColumnTransformer(_transformers, remainder='drop')
_y = _eager_4['churn']
_eager_5 = Pipeline([('preprocessor', _preprocessor), ('model', RandomForestClassifier(n_estimators=20, random_state=42))])
_eager_5.fit(_X, _y)
_est = _eager_5.named_steps.get('model', _eager_5) if hasattr(_eager_5, 'named_steps') else _eager_5
_imp = getattr(_est, 'feature_importances_', None)
if _imp is None:
    _imp = abs(_est.coef_).mean(axis=0) if _est.coef_.ndim > 1 else abs(_est.coef_)
if hasattr(_eager_5, 'named_steps') and 'preprocessor' in _eager_5.named_steps:
    _raw = [str(c) for c in _eager_5[:-1].get_feature_names_out()]
    _names = [next((c[len(p):] for p in ('num__', 'cat__', 'remainder__') if c.startswith(p)), c) for c in _raw]
else:
    _fn = getattr(_est, 'feature_names_in_', None)
    _names = [str(c) for c in _fn] if _fn is not None else [f'feature_{i}' for i in range(len(_imp))]
_eager_6 = pd.DataFrame({'feature_name': _names, 'importance': _imp}).sort_values('importance', ascending=False).reset_index(drop=True)
_eager_6['rank'] = _eager_6.index + 1
_eager_6 = _eager_6.head(2)
df_3 = pl.from_pandas(_eager_6)
df_3.write_csv('feature-importance.csv')

Tips and common mistakes

Load only trusted artifacts. MLflow and joblib loading can execute Python code; path and URI validation do not isolate the model's code.
Use a trained model with tree importance or linear coefficients. An unsupported estimator cannot produce this table merely by adding a model wire.
A high rank is not a quality score for the model. Use Evaluate on held-out predictions as a separate check.
In this sample, the first row is support_calls with importance 0.4; the second is tenure with importance 0.3. Tied features can change order across library versions; do not interpret their ordering as a measured gap.