LIME (Local Interpretable Model-agnostic Explanations)

1) What it is

• LIME = technique to explain individual predictions of any ML model.
• Stands for Local Interpretable Model-agnostic Explanations.
• Idea: Instead of explaining the whole black-box model, approximate it locally around the instance of interest with a simpler interpretable model (like linear regression).

LIME answers: “Why did the model predict THIS for THIS example?”

2) How it works (step by step)

1. Take the instance you want to explain (e.g., a loan application).
2. Create perturbed samples by slightly modifying the input features.
3. Get predictions from the black-box model for these samples.
4. Weight samples by proximity to the original instance (closer samples matter more).
5. Fit a simple, interpretable model (like linear regression or decision tree) on this local neighborhood.
6. Use the coefficients of this simple model to explain which features contributed most.

3) Example

Model predicts: Loan denied

• LIME perturbs features like income, age, debt ratio.
• Black-box predictions on these perturbed points are collected.
• Local linear model shows:
  • Income (-0.4 contribution)
  • High debt (+0.3 contribution)
  • Employment length (+0.1 contribution)

Explanation: Low income + high debt pushed decision toward denial.

4) Strengths of LIME

Model-agnostic → works with any classifier/regressor.
Local focus → explains one prediction at a time.
Human-friendly → produces simple feature-weight explanations.

5) Limitations

Instability: Explanations can change if you perturb data differently.
Local approximation only: Doesn’t guarantee global faithfulness.
Computationally expensive: Requires many perturbations.
Correlated features: Hard to interpret weights when features interact.

6) Python Example

import lime
import lime.lime_tabular
from sklearn.datasets import load_iris
from sklearn.ensemble import RandomForestClassifier

X, y = load_iris(return_X_y=True)
model = RandomForestClassifier().fit(X, y)

explainer = lime.lime_tabular.LimeTabularExplainer(X, feature_names=["f1","f2","f3","f4"], class_names=["setosa","versicolor","virginica"], discretize_continuous=True)

i = 0
exp = explainer.explain_instance(X[i], model.predict_proba, num_features=2)
exp.show_in_notebook()

import lime
import lime.lime_tabular
from sklearn.datasets import load_iris
from sklearn.ensemble import RandomForestClassifier

X, y = load_iris(return_X_y=True)
model = RandomForestClassifier().fit(X, y)

explainer = lime.lime_tabular.LimeTabularExplainer(X, feature_names=["f1","f2","f3","f4"], class_names=["setosa","versicolor","virginica"], discretize_continuous=True)

i = 0
exp = explainer.explain_instance(X[i], model.predict_proba, num_features=2)
exp.show_in_notebook()

7) LIME vs SHAP (quick compare)

Summary

• LIME = explains black-box predictions by approximating the model locally with a simple surrogate.
• Useful for instance-level interpretability.
• Pros: simple, flexible, model-agnostic.
• Cons: unstable, only local, approximation may not reflect true model logic.