Importance Chart
The Importance Chart shows parameter importance as horizontal bars for a selected objective or compares objectives with ALL. Pick the objective selection and sensitivity method, then press Run.
See Sensitivity Analysis for theory and method selection, or the Sensitivity Heatmap for a colored parameter-by-objective grid.
Objective selection and execution
- For multi-objective studies, ALL is the first objective option and the default for newly opened Importance Charts.
- In ALL, one Run computes the selected method separately for every objective. Existing per-objective cached results are reused; only missing entries are computed asynchronously. The UI remains responsive and indicates progress until the request is complete.
- Selecting an individual objective computes and displays only that objective, with bars sorted by descending importance. Single-objective studies keep their existing behavior.
- In constrained studies, Feasible only restricts fitting to feasible trials.
Reading the bars
- ALL groups horizontal bars by parameter, with one bar per objective. Objective colors and a legend identify the bars on one shared horizontal axis.
- Parameter groups are ordered by their maximum importance across objectives, descending. Numeric labels show the method's scores unchanged: there is no extra normalization or combined objective score. Score interpretation still depends on the selected method.
- Numeric axis ticks and subtle vertical gridlines help compare bar lengths. Horizontal separators distinguish parameters (whole parameter groups in ALL), in both light and dark themes.
- Methods that provide R² show it separately for each objective in ALL; methods without R² do not show it.
- Failed or unavailable objective results and unsupported parameters are identified explicitly, never shown as zero. Genuine zero scores remain valid results. Successful results remain visible even if another objective fails or is unavailable.
CSV export
In ALL, CSV uses objective,variable,importance_score,method columns and
exports only successful, available scores. Failed, unavailable or unsupported
results are not exported as numeric zeros. When only part of the results can be
exported, the UI explicitly identifies the export as partial so omissions are
not mistaken for a complete comparison.
Selecting an individual objective keeps the existing CSV schema and behavior.
Numerical-only Spearman and Ridge
Only numerical parameters receive bars or rankings. Nominal categorical parameters, including numeric-looking Optuna choices, are listed separately as Unsupported (categorical). This does not mean zero importance. If every parameter is categorical, the widget shows No supported numerical parameters.
Ridge displays R² (numerical-only): on mixed inputs it assesses only the included numerical features, not excluded categorical effects. All-categorical inputs have no fitted numerical model and no Ridge R². Optuna Int/Float parameters are supported. Spearman excludes non-finite parameter/objective pairs; Ridge excludes rows with any non-finite included numerical feature or objective before sampling and splitting.
Individual-objective CSV export includes unsupported rows with
Unsupported (categorical) in the score field, not a number. These rows are
outside the scored ranking; ALL exports only available scores as described
above.
Cautions
- Scores describe association or predictive contribution on observed data, not causality. Confirm findings with another method.
- Model-based methods depend on fit quality; a poor fit can make rankings unreliable. Low numerical-only Ridge R² does not prove categories irrelevant.
- Other methods have categorical encoding limitations.