AutoML Feature Importance

What it does

This node fits a model on your table and ranks which columns actually drive the target. It handles the preprocessing, trains on a 75/25 split, and measures importance by permutation or from the model’s own scores. Output is the ranking (rank, feature, importance, importance_std, method), plus an optional bar chart on the Artifacts tab. Runs on a kernel.

Inputs

One table: any mix of numeric, string, boolean or categorical feature columns, plus a target column. Rows with a null target are dropped. Classification needs two or more classes, regression a numeric target.

Settings

Data — the feature columns and the target. Task is auto-detected (strings, booleans and low-cardinality integers become classification); set it yourself if the guess is wrong.

Model — which estimator you’re explaining, linear through the tree ensembles to XGBoost and LightGBM. Different models rank differently. Balancing only applies to classification: class weights by default, or over/undersample the training split.

Importance — permutation works for any model and gives you a std, built-in is faster but has neither (and KNN has none at all). The scoring metric is what permutation degrades; more repeats means a steadier ranking and longer runtime. Top N truncates the output.

Visualization — publish the ranked bar chart under a name you pick, or turn it off for table only.