Calculated Results
Results update after calculation and remain above the form.
Detailed Summary
| Minimum | — | Maximum | — |
|---|---|---|---|
| Range | — | Variance | — |
| Feature rows | — | Non-zero rows | — |
| Lowest-ranked feature | — | Top contribution | — |
| Rank | Rank Δ | Feature | Model or Class | Raw | Normalized | Contribution | Cumulative | 95% CI | Direction | Interpretation |
|---|
Interpretation
Selected Feature List
Model Agreement
Formula Used
Normalized Importance = |Feature Importance| ÷ Sum of |All Importances| × 100
Permutation Importance = Baseline Model Score − Permuted Feature Score
Standardized Importance = Coefficient × Feature Standard Deviation
Cumulative Percentage = Running Sum of Absolute Importance ÷ Total Absolute Importance × 100
How to Use
- Select the importance method and machine learning task.
- Enter feature names, model labels, values, and uncertainty.
- Choose normalization, ranking, filtering, and chart settings.
- Press calculate to create the chart and ranked table.
- Review warnings before removing low-importance features.
- Export the chart, data, interpretation, or selected features.
Example Data
| Feature | Model | Importance | Uncertainty |
|---|---|---|---|
| Income | Random Forest | 0.3100 | 0.0200 |
| Credit Score | Random Forest | 0.2800 | 0.0180 |
| Age | Random Forest | 0.2400 | 0.0150 |
| Account Age | Random Forest | 0.1700 | 0.0120 |
Frequently Asked Questions
What does feature importance measure?
It measures how strongly each input contributes to model behavior. The exact meaning depends on the selected importance method.
Does high importance prove causation?
No. Importance explains model reliance, not real-world cause and effect. Confounding variables can influence the ranking.
Why can correlated features receive low importance?
Correlated variables may divide predictive information between themselves. Removing one can increase the measured importance of another.
Why does impurity importance favor some features?
Tree impurity methods can favor continuous or high-cardinality variables. Permutation or SHAP analysis can provide useful comparisons.
Should negative importance values be removed?
Not automatically. Negative permutation values may indicate noise, instability, sampling error, or a metric direction issue.
What cumulative threshold should I use?
Ninety percent is a common exploration setting. Validation performance should decide the final retained feature set.
Can I compare multiple models?
Yes. Enter the same features with different model labels, then use grouped or stacked comparison mode.
What does standardized coefficient importance mean?
It multiplies each coefficient by feature standard deviation. This improves comparisons when feature measurement scales differ.
How are confidence intervals calculated?
The calculator uses the entered uncertainty and a normal critical value. Bootstrap intervals require externally computed uncertainty estimates.
Can selected features replace proper validation?
No. Retrain and evaluate the reduced model using unseen data. Feature selection can otherwise produce optimistic conclusions.