SHAP vs LIME: Explainability Scoring

[SHAP (SHapley Additive exPlanations)] excels at providing theoretically consistent global and local explanations because it is grounded in cooperative game theory. For example, SHAP values guarantee that the sum of feature contributions equals the model's prediction, a property known as efficiency, which is critical for audit-grade compliance. However, this mathematical rigor comes at a cost: exact KernelSHAP computation scales exponentially with the number of features, making it computationally prohibitive for high-dimensional models without optimized implementations like TreeSHAP.

[LIME (Local Interpretable Model-agnostic Explanations)] takes a different approach by approximating the decision boundary locally with an inherently interpretable surrogate model, such as a linear regression. This results in a significant speed advantage, as LIME can generate an explanation in seconds regardless of the underlying model's complexity. The trade-off is fidelity: because LIME samples perturbed instances to learn a local approximation, the explanation can be unstable and may not perfectly reflect the true global behavior of the model, leading to potential inconsistencies in regulated fairness audits.

The key trade-off: If your priority is mathematical consistency and defensibility for regulatory compliance (e.g., EU AI Act), choose SHAP. If you prioritize rapid, model-agnostic prototyping and computational speed across diverse data types like text and images, choose LIME. For enterprise tabular data auditing, SHAP's TreeSHAP variant often nullifies the speed gap, making it the default choice for gradient-boosted models.

SHAP excels at providing theoretically consistent, global and local explanations because it uniquely satisfies three axioms: local accuracy, missingness, and consistency. For example, in high-stakes regulated environments like credit underwriting, SHAP values guarantee that the sum of feature contributions equals the model's prediction, offering a legally defensible audit trail that LIME's local approximations cannot match.

LIME takes a different approach by treating the model as a complete black box, perturbing inputs to learn a simple, interpretable surrogate model locally. This results in significantly faster computation times—often 10x to 100x faster than exact SHAP for complex ensembles—but at the cost of explanation stability. Two runs of LIME on the same instance can produce different explanations, a trade-off that is unacceptable for compliance but often tolerable for exploratory debugging.

The key trade-off: If your priority is mathematical consistency and regulatory defensibility, choose SHAP. Its unified framework provides the same explanation metric for both local and global analysis, which is critical for AI governance platforms requiring audit-ready documentation. If you prioritize computational speed and true model-agnosticism for rapid prototyping across text, image, and tabular data, choose LIME. Consider SHAP when you need to defend decisions to regulators; choose LIME when you need to quickly debug a model before lunch.