Fidelity

A local explanation can be simple, readable, and... totally wrong about the black box! Fidelity measures how closely the explanation reproduces the model behavior it claims to summarize.

For LIME, that means agreement in the local neighborhood. For a surrogate tree, it may mean agreement across a broader dataset.

ℹ

Agreement with the target model

For surrogate $g$ and black box $f$ , local fidelity can be measured with weighted loss:

$Fidelity=1-\frac{\sum_i\pi_x(z_i)L(f(z_i),g(z_i))}{\sum_i\pi_x(z_i)}$

The metric must use the same scope claimed by the explanation.

Compare Black-Box and Surrogate Decisions Point by Point

The star is the loan applicant the explanation is written for. Each dot is a nearby applicant. Drag the radius to change how far 'nearby' reaches, and raise the surrogate's terms to let its straight-line rule bend toward the model's real, curved boundary.

Evaluation radius0.30Surrogate terms1

Weighted fidelity

0.89

Neighbors in scope

6 of 18

Disagreements in scope

3 more disagreements sit outside the radius — invisible to this fidelity score, but real if anyone applies the explanation out there.

black box and surrogate agree black box and surrogate disagree outside the evaluation radius true (black-box) boundary

Fidelity belongs to a scope. Widen the radius and the same surrogate can start missing the neighbors that matter.

High fidelity does not guarantee usefulness. A 200-rule surrogate can mimic the model and fail every human reader. Low fidelity makes the explanation factually weak regardless of presentation quality.

Report fidelity alongside complexity and the evaluation region.

Checkpoint

Can a globally low-fidelity surrogate still be useful?

←PreviousXAI in Agentic AI SystemsLarge Language Model Explanations Next→FaithfulnessEvaluating Explanations