Interpretation

What an IMV value says

IMV is the relative change in equivalent-coin weight from the declared baseline to the declared enhanced predictor. For example, 0.10 means the enhanced weight is 10% larger than the baseline weight. It does not mean ten percentage points of accuracy, ten percent of outcome variance, or ten percent of mutual information.

Always retain the arrow in prose: “IMV from the intercept-only model to model X” is less ambiguous than “model X has IMV 0.10.”

Baseline dependence

The denominator is the baseline weight. Two studies using different baselines can obtain different IMV magnitudes for the same enhanced predictions. Compare values only when the outcome definition, evaluation rows, and baseline are compatible.

A constant prevalence model is a useful null baseline when it is estimated only on training data. A domain model, an earlier production model, or an ablated architecture can also be valid if it matches the scientific contrast.

Negative values

A negative IMV means the enhanced predictor assigned less geometric mean likelihood to held-out outcomes than the baseline. Possible causes include overfitting, harmful features or components, poor optimization, probability miscalibration, distribution shift, or sampling variability.

For SHAP-IMV, a negative feature attribution means the feature reduced the global coalition value on average. It does not mean the feature predicts the negative outcome class.

Below-chance results

A substantially below-0.5 geometric mean likelihood has no equivalent coin and therefore no IMV weight. Treat the resulting warning and NaN as diagnostic information, not a value to replace with zero. Inspect calibration and data alignment, and report information_deficit when the severity itself matters.

Matrix semantics

Output	Row role	Column role	Symmetry
AblationIMV.calculate_imv_matrix	Enhanced model	Baseline model	Directional; generally neither symmetric nor antisymmetric
MulticlassIMV.multinominal_imv_matrix	First class in a pair	Second class in a pair	Symmetric by construction

Read an ablation matrix down one baseline column when comparing variants against that fixed baseline. Do not rank cells across columns as if their denominators were the same.

What IMV is not

• IMV is not mutual information, even though it uses an entropy-shaped inverse.
• IMV is not itself a calibration score; it is sensitive to calibration because it consumes probabilities.
• IMV is not defined by hard class predictions or accuracy.
• This implementation is not a regression metric.
• Fold or seed dispersion is not automatically a standard error or confidence interval.