Local explanations

A local explanation approximates model behaviour around one selected input and a defined perturbation neighbourhood.

It can answer questions such as:

• Which input components were associated with the largest local output changes?
• Did those components increase or decrease the surrogate output?
• How well did the surrogate approximate sampled model behaviour?

It does not automatically answer:

• Which features are globally important across the dataset?
• What causal mechanism produced the prediction?
• What internal reasoning process the model used?
• Whether the prediction or explanation is fair, safe, or correct?

Local explanations should be compared across nearby inputs, seeds, perturbation settings, and relevant subgroups before strong conclusions are drawn.