Pick a point on the decision space, and watch LIME build a local explanation:
perturb, query the black box, fit a weighted linear model, read off which features mattered.
Black-box model
This is the "unknown" classifier LIME tries to explain locally.
LIME parameters
Black-box prediction–
LIME local R²–
x₁ weight–
x₂ weight–
How it works
1Pick a point to explain
2Perturb it — create N noisy neighbors
3Query the black box on each neighbor
4Weight neighbors by proximity (kernel)
5Fit a weighted linear model
6Read the coefficients — that's the explanation
ξ(x) = argming∈G L(f, g, πx) + Ω(g)
Where πx(z) = exp(−d(x,z)² / σ²)
and g is a local linear model.
Decision space — click to pick a point
Background color = black-box prediction (blue vs red).
Click anywhere to select the point LIME will explain.
LIME perturbation cloud
Dots = perturbed samples. Color = black-box label.
Opacity = kernel weight (closer = brighter).
Dashed line = local linear boundary from LIME.
Feature importance (LIME weights)
The linear model's coefficients β₁ (x₁) and β₂ (x₂).
Positive = pushes toward class 1 (red), negative = pushes toward class 0 (blue).
Longer bar = stronger local influence.