Local Interpretable Model-agnostic Explanations (LIME)

LIME operates on the principle of black-box access, requiring no knowledge of the model's internal weights, gradients, or architecture. It treats the original model as a queryable oracle, sending perturbed inputs and receiving prediction probabilities in return. This property makes LIME universally applicable across logistic regression, gradient-boosted trees, deep neural networks, and even proprietary APIs. The explanation system only needs the ability to call model.predict(), making it a truly post-hoc explanation method that can be retrofitted onto any existing production system without modifying the underlying model.

Rather than attempting to explain the entire global decision boundary, LIME enforces local fidelity by fitting an interpretable surrogate model to a small neighborhood around the instance of interest. The exponential kernel assigns higher weights to perturbed samples closer to the original instance, ensuring the explanation accurately reflects the model's behavior in that specific region. This is measured by the fidelity-interpretability trade-off: a simple linear model may not perfectly capture the local curvature of a deep network, but it provides a human-understandable approximation that is faithful where it matters most.

LIME decouples the model's native feature space from the explanation space through an interpretable representation. For text, this is a bag-of-words binary vector indicating word presence. For images, superpixel segmentation groups contiguous, perceptually similar pixels into atomic regions. For tabular data, feature discretization bins continuous values into human-friendly categories. The surrogate model is trained on these interpretable features, and the resulting feature importance scores are mapped back to the original input, ensuring explanations use concepts humans can reason about.

The local neighborhood is constructed through perturbation sampling, which varies by data modality:

• Text: Randomly removes words or tokens from the instance to create counterfactual documents
• Images: Turns superpixels on or off by masking them to a neutral color
• Tabular: Draws samples from a normal distribution centered on each feature value Each perturbed sample is weighted by its cosine distance (for text) or Euclidean distance (for tabular/image) to the original instance, enforcing the locality constraint.

To prevent cognitive overload, LIME employs Lasso regression (L1-regularized linear model) as the default surrogate. The L1 penalty drives the coefficients of less important features to exactly zero, producing a sparse linear model that highlights only the top-K most influential features. This explanation regularization ensures the output is concise—typically 5-10 features—making it suitable for direct human consumption. The number of features can be tuned as a hyperparameter to balance completeness against clarity.

Because perturbation sampling is inherently stochastic, explanation stability is a critical property. Multiple runs with different random seeds should yield consistent feature rankings. The kernel width hyperparameter controls this stability: a wider kernel averages over more samples (higher stability, lower local fidelity), while a narrower kernel tightly focuses on the instance (higher fidelity, lower stability). Frameworks like OptiLIME and Bayesian LIME address this by automatically tuning the kernel width and providing uncertainty estimates alongside feature importance scores.