LIME: Local Interpretable Model-agnostic Explanations
What You'll Learn
You'll understand how LIME explains individual predictions from black-box models by creating simple, interpretable approximations around specific data points.
The Core Philosophy
LIME (Local Interpretable Model-agnostic Explanations) operates on a brilliant insight: even if your overall model is incredibly complex—a deep neural network, a massive ensemble, or any "black box"—the decision boundary near one specific prediction might be surprisingly simple.
Think of it like explaining a mountain range. Globally, the terrain is complex and jagged. But if you zoom into one small patch where you're standing, it might look almost flat—approximated well by a simple plane.
How LIME Works Conceptually
For a single prediction you want to explain, LIME:
- Generates perturbed samples around that instance (neighbors in feature space)
- Gets predictions from the black-box model for these neighbors
- Fits a simple, interpretable model (typically linear regression) using these samples, weighted by proximity to the original instance
- Uses the simple model's coefficients as feature importance for that specific prediction
The "local" part means it only explains this one prediction, not the entire model. The "model-agnostic" part means it works with any model—you only need access to predictions, not internal parameters.
Why Local Linear Approximations?
Complex models create non-linear, high-dimensional decision boundaries. But LIME asks: "What's the simplest explanation for this particular decision?" A local linear approximation is interpretable—you can understand which features pushed the prediction up or down—while being flexible enough to approximate most smooth functions in a small neighborhood.
Key Takeaway: LIME explains individual predictions by fitting an interpretable linear model in the neighborhood of a specific instance, trading global accuracy for local fidelity and interpretability.