Sensitivity Heatmap

The Sensitivity Heatmap widget shows one importance method's scores for every parameter against every objective at once, laid out as a colored grid: one row per parameter, one column per objective. Where the Importance Chart ranks parameters for a single objective, the heatmap trades that detail for breadth — a single glance tells you which parameters matter for which objectives across a multi-objective study, and which parameters are irrelevant everywhere.

The scores themselves come from the same sensitivity analysis methods as the Importance Chart; this page covers only the widget. For what each method measures and how to choose one, see Sensitivity Analysis.

Method Selection and Execution

1. Pick the method from the dropdown. It is grouped by family so each method's character is clear:
   • Correlation / Linear — Spearman, Ridge
   • Tree-based — RF-ANOVA, MDI, SHAP, Permutation
   • Global Sensitivity — Sobol (first-order and total-effect)
2. Low-cost methods (Spearman / Ridge) compute automatically when selected. The heavier methods wait for the Run button, so switching to an expensive method does not trigger a long computation by accident.
3. On a constrained study, toggle Feasible only to fit the underlying model on feasible trials alone.

Reading the grid

• Rows are parameters, columns are objectives. Read down a column to see which parameters drive that one objective; read across a row to see whether a parameter matters for one objective, several, or none.
• Each cell shows the numeric score (2 decimals) on top of its color, so you can compare exact values as well as the visual pattern.
• Color depends on the method's sign:
  • Signed methods (Spearman, Ridge) use a diverging color scale over $[-1, 1]$, so the sign — a parameter that pushes an objective up versus down — is visible, not just the magnitude.
  • Non-negative methods (tree-based and Sobol) are normalized per column by that column's maximum and shown on a sequential scale. Colors are therefore comparable within a column (parameters for one objective), but not directly across columns — the brightest cell of each objective is just that objective's most important parameter.

Sensitivity Heatmap vs. Importance Chart

	Importance Chart	Sensitivity Heatmap (this widget)
Scope	one objective at a time	all parameters × all objectives at once
Layout	ranked bars	parameter-by-objective color grid
Best for	reading the exact ranking for one objective	spotting cross-objective patterns and dead parameters
Method	the same importance methods	the same importance methods

Use the heatmap to find where to look, then open the Importance Chart on a specific objective for the precise ranking.

Caveats

• Off-diagonal comparisons of non-negative methods can mislead. Because each column is normalized on its own, a bright cell under objective A and a bright cell under objective B do not represent the same absolute importance. Compare magnitudes only within a column.
• Importance scores measure association / predictive contribution on the observed data, not causation. As with the Importance Chart, confirm findings with a second method before drawing conclusions.
• Model-based methods (tree-based, Sobol) depend on the quality of the fitted model; a poor fit makes every column unreliable.