Perturbation stability
availableSlightly perturbs numerical observations and measures how consistently the analytical result is recovered.
Features
Stability analysis
availableTest whether an analytical result persists when the underlying data or analytical conditions are deliberately perturbed.
Overview
Stability analysis examines whether an analytical result remains similar when the conditions that produced it are changed in controlled ways.
This provides a complementary view to ordinary performance or separation metrics by testing sensitivity rather than evaluating only the original run.
Stability modes
Different stability modes test different sources of sensitivity.
Available
Planned
Algorithmic stability
coming soonTests sensitivity to algorithmic randomness, such as alternative model initializations.
Noise stability
coming soonTests whether irrelevant or deliberately noisy features materially alter the result.
Subsampling stability
coming soonMeasures whether structure persists when analyses are repeated on subsets of observations.
Outputs
- Run-level stability metrics
- Per-cluster stability information where applicable
- Integration with the Stability evaluation pillar
- Comparison against other runs and configurations
Interpretation
- High stability does not by itself imply that an analytical structure is meaningful.
- Different stability tests probe different failure modes and should not be treated as interchangeable.
- Sensitivity may reflect properties of the dataset, preprocessing, model, parameters, or representation.