Uncertainty-aware XAI (robustness)

[breimanntools]

Problem

Predictions and explanations lack confidence estimation.

Goal

Quantify uncertainty in predictions and feature importance.

Tasks

• Implement model ensembles
• Compute SHAP variance across models
• Add Bayesian uncertainty (PyMC / ArviZ)
• Identify stable vs unstable features

How this improves AAanalysis

• Distinguishes:
  robust biological signals vs noise
• Prevents misleading interpretations
• Guides experimental prioritization

Acceptance criteria

• Uncertainty provided per prediction and feature