Classical ML, validation, metrics, interpretation, debugging · insight-0179

Group correlated features before reading SHAP rankings

Situation
Several features measure the same thing (fare level, average fare, revenue per passenger) and SHAP ranks them separately.
Action
Cluster correlated features, sum SHAP values within each group for global importance, and state that within-group splits are not meaningful.
Reason
When features are redundant, the model can use any of them and attributions spread across them in ways that depend on training randomness.
Boundary
Global SHAP summaries, feature-pruning decisions and executive explanations.

Symptom or risk

Credit is split among redundant features, so each looks modest and a weaker but unique feature appears more important.

Local explanations of a single prediction may still be useful at feature level if the audience understands redundancy.

Groups are easier to explain but hide which column to monitor; keep the mapping.

Dropping a feature because of low SHAP importance when its twin carries the signal.