LIME vs. SHAP: Choosing the Right Explainable AI Method for Scientific Research

Choosing between LIME and SHAP for Explainable AI in scientific research involves a trade-off between speed and robustness. LIME offers fast, local explanations but can be unstable, while SHAP provides consistent, global insights at a higher computational cost.

Yasmin Haddad

November 1, 2026 · 6 min read

For researchers who need to interpret complex neural network decisions, the choice between LIME and SHAP comes down to one central trade-off: speed versus robustness. LIME delivers fast, local explanations for individual predictions by fitting a simple surrogate model around a single instance, but those explanations can be unstable across repeated tests. SHAP, grounded in cooperative game theory, produces consistent feature attributions that work both locally and globally, at a higher computational cost. A comparative study presented at an IEEE conference frames the decision exactly this way: LIME excels at quick local explanations that are easy to implement, while SHAP offers more robust and global insights despite its computational overhead.

Introduction to Explainable AI: LIME and SHAP

Complex neural networks achieve strong predictive performance, but their internal decision logic is opaque to the scientists who rely on them. Explainable AI (XAI) methods address this gap by attaching interpretable outputs to otherwise black-box models. Two techniques dominate practical use in scientific research: LIME (Local Interpretable Model-agnostic Explanations) and SHAP (SHapley Additive exPlanations). Both are post-hoc and model-agnostic, meaning they can be applied to any trained machine learning model without modifying it, as a perspective paper on the two methods notes.

The distinction that matters most for researchers is scope. Local interpretability explains why a model made one specific prediction for one specific instance. Global interpretability characterizes how the model behaves across an entire dataset. LIME is fundamentally a local method; SHAP spans both. This difference in scope, together with a difference in stability and cost, drives the entire selection decision.

Both methods have proven their value in applied scientific settings. A study in the electronics research literature describes how SHAP and LIME were applied together to provide global and local explanations of model predictions in wastewater treatment research, where machine learning models such as artificial neural networks, gradient boosting machines, and random forests predicted variables including BOD and TSS from multi-year operational data. The interpretability layer revealed the most influential features driving those predictions, turning an opaque predictor into a source of scientific insight.

LIME: Local Interpretability Through Surrogate Models

LIME answers a narrow question: why did the model produce this particular output for this particular input? Its core mechanism is local approximation. Rather than trying to explain the entire complex model, LIME approximates the model's behavior in the neighborhood of a single instance using an interpretable surrogate—typically a simple, locally fitted model. Because the approximation is performed for a specific instance, each explanation is self-contained and scoped to that one prediction.

This design has practical consequences. Because the surrogate is simple and the analysis is confined to one instance's neighborhood, LIME is relatively easy to implement and fast to run. The IEEE comparative study highlights this as LIME's principal strength: it excels at performing local explanations through interpretable models, making it comparatively easy to put into practice. A perspective paper on SHAP and LIME reinforces the point, observing that LIME is much faster than SHAP, especially with tree-based models.

The same mechanism, however, introduces a weakness. LIME's approximation process relies on perturbing the instance being explained—generating variations of the input and observing how the model responds. The comparative study reports that LIME explanations are not very stable across tests due to these instance perturbations. In practice, that means running LIME twice on the same prediction can yield noticeably different feature importance rankings, a serious concern when an explanation is meant to support a scientific claim.

There is also a structural limitation noted in the SHAP-and-LIME perspective paper: because LIME fits a local linear model, it fails to capture non-linear associations, whereas SHAP may detect such associations depending on the model used. For researchers whose neural networks learn strongly non-linear relationships, this matters.

LIME vs. SHAP: A Decision Framework for Researchers

The comparison below juxtaposes the two methods against a consistent set of criteria—their underlying theory, the scope of their explanations, their stability, their computational cost, and their primary use case. The pattern that emerges is a clean trade-off: LIME trades robustness for speed, while SHAP trades speed for consistency and global reach.

LIME and SHAP compared across theory, scope, consistency, cost, and use case
Criterion	LIME	SHAP
Core principle	Approximates the complex model locally with an interpretable surrogate model, fitted for a specific instance.	Grounded in game theory; attributes feature importance using Shapley values that fairly distribute the prediction among features.
Scope of explanation	Local only—explains how the model works for a specific instance; one plot per instance.	Both local and global—supports understanding of individual predictions and overall model behavior.
Consistency	Not very stable across tests, due to instance perturbations in the approximation process.	Consistent explanations; the game-theoretic foundation ensures feature attributions are reliable.
Computational cost	Relatively low; faster for generating individual explanations, especially compared to SHAP on tree-based models.	Higher; cost increases with the number of features and instances due to the permutation-based approach.
Primary use case	Quick local checks on single predictions when implementation simplicity and speed matter most.	Robust local and global analysis when explanation reliability and model-wide insight are required.

LIME and SHAP compared across theory, scope, consistency, cost, and use case

Criterion

LIME

SHAP

Core principle

Approximates the complex model locally with an interpretable surrogate model, fitted for a specific instance.

Grounded in game theory; attributes feature importance using Shapley values that fairly distribute the prediction among features.

Scope of explanation

Local only—explains how the model works for a specific instance; one plot per instance.

Both local and global—supports understanding of individual predictions and overall model behavior.

Consistency

Not very stable across tests, due to instance perturbations in the approximation process.

Consistent explanations; the game-theoretic foundation ensures feature attributions are reliable.

Computational cost

Relatively low; faster for generating individual explanations, especially compared to SHAP on tree-based models.

Higher; cost increases with the number of features and instances due to the permutation-based approach.

Primary use case

Quick local checks on single predictions when implementation simplicity and speed matter most.

Robust local and global analysis when explanation reliability and model-wide insight are required.

Read vertically, the table shows why the methods diverge: every row traces back to the same root difference. LIME's local surrogate produces speed but instability and local-only scope; SHAP's Shapley computation produces consistency and dual scope but higher cost. Read horizontally, it gives a verdict: if the priority is speed for individual predictions, LIME aligns with the goal; if the priority is robust, consistent explanations across the model, SHAP does.