Course contentsShow
← Machine Learning and Deep Learning
Lesson 3229 of 3,53870. Interpretability: Feature-Based MethodsFree lesson

LIME Stability and Reliability Issues

Understanding limitations including explanation instability across repeated runs and perturbation sensitivity.

LIME Stability and Reliability Issues

What you'll learn: Why LIME explanations can vary unpredictably across repeated runs and different perturbation strategies.

The Core Problem

LIME generates explanations by randomly perturbing input samples and fitting a local linear model. However, this randomness introduces instability—run LIME twice on the same prediction, and you may get noticeably different explanations. This undermines trust: if explanations change each time, which one is "correct"?

Sources of Instability

Perturbation Sensitivity

The random samples LIME generates around your data point strongly influence the surrogate model. Different perturbation sets can emphasize different features, especially in high-dimensional spaces where the neighborhood is sparsely populated.

Local Neighborhood Definition

LIME's kernel width determines what counts as "local." Too narrow, and you have few samples to fit a reliable model. Too wide, and you're no longer truly local—the linear approximation becomes poor.

Model Fitting Variance

Even with the same perturbations, the weighted linear regression can be sensitive to outliers or specific sample configurations, leading to coefficient instability.

Practical Implications

Think of LIME like asking directions from locals who each take slightly different routes—the general direction might be consistent, but specific street names vary. For critical applications (medical diagnosis, loan decisions), this unreliability is problematic. You can't confidently say "this feature caused the prediction" when re-running produces a different top feature.

Key Takeaway: LIME's reliance on random perturbations and local approximations makes explanations unstable across runs and sensitive to hyperparameters—always validate consistency by running multiple times and consider ensemble approaches when stability matters for your application.