LIME or SHAP? A Researcher's Decision Framework for Model Explainability

Computational scientists choosing between LIME and SHAP for model explainability must consider their specific needs regarding prediction scope, model architecture, and computational resources. This article provides a decision framework to help researchers select the most appropriate technique, weighing factors like local interpretability, global consistency, and efficiency trade-offs.

Yasmin Haddad

October 29, 2026 · 6 min read

Computational scientists who need to explain a complex machine learning model face a deceptively simple choice between two dominant techniques: LIME and SHAP. The short answer is that the right pick depends on three questions. Do you need to explain one prediction or the model's overall behavior? Is your model a tree ensemble? And how much compute can you afford? LIME delivers fast, local explanations for essentially any black-box model but produces unstable results across repeated runs, according to a comparative analysis of the two methods published in IEEE. SHAP, grounded in cooperative game theory, provides consistent local and global feature attributions—but exact computation is exponential in the number of features, unless you use a specialized variant such as TreeSHAP, which runs in polynomial time for tree-based models.

LIME: Local Interpretability Through Surrogate Models

LIME—Local Interpretable Model-agnostic Explanations—answers a narrow question: why did the model make this prediction for this instance? Rather than trying to open the black box, it builds a small, interpretable surrogate model—typically a linear approximation—around the specific point being explained. The method perturbs the instance, observes how the complex model responds, and fits the simple model to that local behavior. The surrogate's coefficients then serve as the feature explanation.

Because it only queries the model as an oracle, LIME works with any algorithm. In a systematic review of LIME and SHAP applications in Alzheimer's disease detection published in PMC, researchers noted that LIME, "a model-agnostic approach, provided local explanations for AD classification on MRI images"—no retraining or architectural knowledge required. This flexibility is LIME's core practical asset for scientists working with heterogeneous models, from imaging pipelines to genomics classifiers.

The cost is two-fold. First, because the surrogate is local and linear, it cannot account for non-linear dependencies among features locally—a limitation explicitly noted in a perspective on SHAP and LIME posted on arXiv. Second, because explanations depend on random instance perturbations, LIME's output is not very stable across repeated tests; the same prediction can yield different feature rankings on each run. That instability matters in scientific settings where reproducibility of the explanation itself is part of the claim.

SHAP: Game Theory for Unified Feature Attribution

SHAP—SHapley Additive exPlanations—takes a principled route borrowed from cooperative game theory. It treats each feature as a "player" and computes its Shapley value: the fair contribution of that feature to a prediction, averaged over all possible orderings and combinations of features. In game-theoretic terms, the Shapley value is the unique allocation that satisfies certain fairness axioms, which is why SHAP is often described as a unified framework for feature attribution. Every prediction is decomposed as an additive sum of per-feature contributions that sum to the model's output.

Where LIME is local by construction, SHAP offers both. A review of LIME and SHAP in Alzheimer's detection puts it plainly: "While LIME focuses on local interpretability, SHAP offers global and local insights with dual interpretability." Individual Shapley values explain single predictions, while aggregating values across a dataset reveals which features drive the model's overall behavior. An urban exposure-mapping study reviewed in Electronics (MDPI) illustrates this dual use: using SHAP, the study identified key spatial and demographic factors influencing EMF intensity, supporting transparent, data-driven mapping for urban planning and public health assessment.

The consistency advantage is real. A comparative study published on IEEE Xplore found that SHAP provides consistent explanations because its Shapley values "ensure fair feature contributions across the dataset," whereas LIME's perturbation-based approach does not. But that theoretical soundness carries computational weight, which is where SHAP's variants enter the picture.

Decision Framework: LIME vs. SHAP for Scientific Research

The framework below maps the three documented trade-offs—explanation scope, model architecture, and computational budget—onto concrete recommendations. Start with your model class, then your explanation goal, then your resources.

Decision matrix for selecting an XAI technique, based on documented mechanisms and trade-offs
Decision point	If yes…	If no…
Is your model a tree ensemble (XGBoost, LightGBM, random forest, etc.)?	Use TreeSHAP: exact Shapley values in polynomial time via the tree structure.	Proceed to the next question.
Do you need global explanations of overall model behavior, or stable, consistent attributions across the dataset?	Use SHAP (TreeSHAP for trees; KernelSHAP for other models, accepting exponential-time approximation).	Proceed to the next question.
Are computational resources limited, or do you need quick, local explanations of individual predictions on any black-box model?	Use LIME: fast, easy to implement, model-agnostic—but verify stability across repeated runs.	Default to SHAP for consistency and theoretical grounding.

Decision matrix for selecting an XAI technique, based on documented mechanisms and trade-offs

Decision point

If yes…

If no…

Is your model a tree ensemble (XGBoost, LightGBM, random forest, etc.)?

Use TreeSHAP: exact Shapley values in polynomial time via the tree structure.

Proceed to the next question.

Do you need global explanations of overall model behavior, or stable, consistent attributions across the dataset?

Use SHAP (TreeSHAP for trees; KernelSHAP for other models, accepting exponential-time approximation).

Proceed to the next question.

Are computational resources limited, or do you need quick, local explanations of individual predictions on any black-box model?

Use LIME: fast, easy to implement, model-agnostic—but verify stability across repeated runs.

Default to SHAP for consistency and theoretical grounding.

The reasoning behind each cell: the tree question comes first because it changes SHAP's cost profile entirely—TreeSHAP's polynomial-time computation removes the main argument against SHAP for tree ensembles, as documented by the AAAI analysis and the treeshap implementation. The scope question follows because it is the hardest constraint to work around: LIME has no native global view, so whole-model interpretation effectively requires SHAP. The resource question resolves the remaining cases; if only local, approximate insight is needed and compute is tight, LIME's speed and model-agnosticism win, provided the instability the IEEE study documented is acceptable—or is checked by running the explanation multiple times.

One worked illustration: a researcher explaining an XGBoost mortality classifier who needs both per-patient explanations and a dataset-level feature summary should choose TreeSHAP outright. A researcher explaining a deep hybrid model on a single batch of predictions, with limited compute and no tree structure available, faces KernelSHAP's exponential-time exact computation versus LIME's faster but unstable local approximations—a genuine trade with no free option.