Why Do SHAP and LIME Disagree on the Same Model?
Direct Answer
SHAP and LIME use fundamentally different mathematical approaches to approximate model explanations. SHAP is based on Shapley values from cooperative game theory and provides globally consistent attributions. LIME fits a local linear model around a single prediction. When the model's decision boundary is non-linear or unstable, these two approximations will diverge — and neither is "correct."
The disagreement between SHAP and LIME is not a bug — it is information. It reveals something important about the model's internal structure.
**SHAP (SHapley Additive exPlanations):**
Computes the marginal contribution of each feature across all possible feature combinations. This is computationally expensive but globally consistent — the same feature will get approximately the same SHAP value across similar instances.
**LIME (Local Interpretable Model-agnostic Explanations):**
Generates synthetic neighbourhood samples around a prediction point, then fits a simple linear model to those samples. The explanation is valid only locally and can vary significantly between nearby predictions.
When these methods disagree, it means the model's decision surface in the neighbourhood of that prediction is non-linear, or the model has learned unstable feature interactions.
**Real-world example:** A credit model assigning a 68% default probability to a self-employed applicant. SHAP assigns +0.23 importance to "income_volatility" (increases default risk). LIME assigns -0.18 importance to the same feature (decreases default risk). This Directional Conflict indicates the model behaves inconsistently for self-employed applicants — a finding that would not appear in any aggregate accuracy metric.
Full Article Access
Continue reading
Enter your details to unlock the remaining 2 sections — free access, no obligation.
NA
Written by
Nodex8 AI Research
AI Research Team
The research team of Nodex8 AI focuses on global AI governance agenda, policy to code maturity across the globe, theoretical and empirical explainable AI research and technology advancement in the domain.
View research publications