LIME and SHAP are two widely used methods for explaining why a machine learning model made a specific prediction, rather than leaving that decision as an unexplainable black box. Both work by estimating how much each individual input feature contributed to a particular output, giving a human-readable breakdown of what actually drove the model's answer.

‍

‍Why it matters

Complex models, especially deep learning ones, are often accurate but nearly impossible to interpret directly, which becomes a real problem when a decision needs to be explained to a regulator, a customer, or an internal reviewer trying to catch a mistake. If a loan application gets rejected by a model, the model said so isn't an acceptable answer in a lot of regulated industries, someone needs to be able to point to which factors actually drove that outcome. LIME and SHAP exist specifically to make that explanation possible without needing to redesign the underlying model itself.

‍

‍How teams use them

Teams typically use these tools to spot-check model behavior, running SHAP across a batch of predictions to look for a feature the model is weighting more heavily than it should, like an outcome that appears to lean on a characteristic it shouldn't be using at all. Both are available as open source Python libraries, which is a big part of why they became the default choice for model explainability work rather than something teams build from scratch.

‍

‍How the two methods differ

LIME works by approximating a complex model locally with a simpler, interpretable model built just around one specific prediction. SHAP is based on game theory, specifically Shapley values, and calculates each feature's contribution more rigorously, though at a higher computational cost. Neither one is strictly better, LIME tends to be faster for a quick check, while SHAP is generally considered more theoretically consistent when the stakes are higher.