Stability and Robustness

Two nearly identical loan applications should not receive wildly different explanations unless the model itself crosses a meaningful boundary.

Stability measures explanation consistency under small input or sampling changes. Robustness asks whether the method resists intentional manipulation.

Local explanation sensitivity

A simple stability score compares explanation vectors:

Stability(x,x)=1E(x)E(x)xx+ϵStability(x,x')=1-\frac{\|E(x)-E(x')\|}{\|x-x'\|+\epsilon}

The exact metric depends on the explanation type, but the principle is consistent: small benign changes should not create arbitrary stories.

Compare Neighboring Explanations

Add a benign input perturbation and rerun the stochastic explainer with a new seed. Compare coefficient rank, sign, and distance.

Explanation distance

0.284

Original

feature 10.420
feature 2-0.310
feature 30.180
feature 40.090

Neighbor

feature 10.455
feature 2-0.307
feature 30.082
feature 4-0.058
Stability requires context. Large explanation changes are expected when the model itself crosses a real boundary.

Run stochastic methods with multiple seeds. Test paraphrases for text, small transformations for images, and plausible neighboring records for tabular data.

Also test adversarial cases. Explanation methods can be manipulated to hide bias while preserving model predictions.

Checkpoint

When is explanation instability expected rather than defective?