LIME — Local Interpretable Model-Agnostic Explanations

Click any point in the decision space to generate a local linear explanation — see how complex model decisions can be approximated locally

Model & Data

Model

Dataset

LIME Parameters

Neighborhood Radius

r 0.30

Smaller radius = more local; larger = more global

Neighbor Samples

n 150

More samples = more stable explanation

Explain a Point

Click anywhere on the decision boundary

Selected point — click to select —

What's happening?

Click any point on the decision boundary to run LIME and see why the model makes that prediction at that location.

Key Concepts ▾

What is LIME? LIME fits a simple interpretable model (linear) locally around a prediction — it explains what the complex model is doing near that specific point, not globally.

Local linearity: complex models are often locally linear — zoom in far enough and any smooth boundary looks like a line. LIME exploits this by fitting a line only in the small neighborhood of interest.

Neighborhood radius: too small = too few neighbors, noisy explanation. Too large = the linear approximation poorly captures the nonlinear boundary. Tune for your problem.

Model-agnostic: LIME only needs predict(x) — it works on any black-box model: neural nets, random forests, SVMs, even remote APIs. It never looks inside the model.

Interpreting weights: positive weight = feature pushes toward class 1. Negative weight = feature pushes toward class 0. Magnitude = strength of influence at this specific point.

Decision Boundary — Neural Net on Moons dataset

Model → Class 0 Model → Class 1 Train class 0 Train class 1 LIME neighbors Local boundary

LIME Feature Importance — local linear approximation

Selected point in decision space

Feature weights (push toward class →)

Pushes toward class 1 Pushes toward class 0

Select a point on the Decision Boundary tab, then switch here to see the LIME feature importance explanation.

How LIME Works — 5-step algorithm

Select a point to explain

Pick any input point x* whose prediction you want to understand. This is the point whose local neighborhood LIME will explore.

Sample neighbors in the neighborhood

Randomly sample points z in the neighborhood of x* (within radius r), then query the black-box model f(z) to get their predicted classes.

Weight neighbors by proximity

Assign weight πₓ*(z) = exp(−d(x*,z)²/σ²) to each neighbor — points closer to x* get higher weight. This ensures the local model focuses on the immediate region.

Fit a weighted linear model

Minimize ξ(x) = argmin_{g∈G} L(f, g, πₓ*) + Ω(g) — find the simplest linear model g that best approximates f locally. The weights πₓ* ensure fidelity near x*.

Coefficients = the explanation

The linear model coefficients w₁, w₂ are the LIME explanation. w₁ > 0 means feature x₁ pushes toward class 1; w₁ < 0 means it pushes toward class 0. Magnitude shows importance.

Selected Point

x₁ —

x₂ —

Class —

Confidence —

LIME Explanation

x₁

—

x₂

—

→ class 1 → class 0

Local Model Fit

—

R² of local linear model

R² near 1.0 = local model fits well. Low R² = neighborhood too large or boundary too curved.

Neighborhood

0.30

radius r

150

samples n