Feature Importance
Rank variables by their predictive importance for a target using gradient boosting and permutation importance.
Overview
Feature Importance measures how much each variable contributes to predicting a target. It fits a gradient boosting model and reports both built-in importance scores and permutation importance across multiple quantiles (Q10, Q25, Q50, Q75, Q90), giving you a sense of each variable's impact and the stability of that estimate.
The dataset must have at least 5 epochs to run feature importance.
How to Run
- Open Feature Importance from the navigation.
- Select Modeling Combinations — choose one or more Country/Region/Grouping triplets.
- Select Targets — the variables you want to predict (defaults to the Sales group).
- Select Variables — the candidate predictors (defaults to all variables except Sales and Events).
- Click Get Feature Importance.
Reading the Results
Table
The results table shows one row per combination × target × variable with columns for each quantile (Q10–Q90). It is sorted by Q50 descending by default.
You can:
- Filter by combination, target, or variable group.
- Sort any column.
- Toggle columns with the Show/Hide Columns button.
Plots
Plots are organized into tabs by variable group (Media, Price, Distribution, etc.). Each tab contains scatter plots — one per target — showing the quantile spread for every variable.
Create Model Shortcut
Select rows in the table and click Create Model to jump directly into model building with those variables pre-selected. This is a convenient way to go from exploration to modeling.
Tips
- A high Q50 with a wide Q10–Q90 spread means the variable is important but its ranking is unstable — investigate further.
- Compare importance across multiple targets to find variables that are broadly predictive vs. target-specific.
- Use this as a pre-modeling step to reduce the variable set before building a model.
See Also
- Correlation — measure pairwise relationships between variables
- Models — build and manage models