LIME (Local Interpretable Model-agnostic Explanations)

LIME does not attempt to explain the entire global behavior of a complex model. Instead, it focuses on local fidelity—explaining why a specific prediction was made for a single instance. It generates a new dataset of perturbed samples around the instance of interest, weights them by their proximity to the original instance, and trains an inherently interpretable model like a sparse linear regression or a shallow decision tree on this local neighborhood. This local surrogate faithfully represents the black-box's decision boundary in that specific region, providing a high-fidelity explanation that would be impossible to capture with a single global approximation.

A defining property of LIME is its complete model-agnosticism. The algorithm treats the underlying classifier as a total black box, requiring no access to its internal weights, gradients, or architecture. It operates solely by observing the relationship between inputs and outputs. This means a single LIME implementation can explain a random forest, a gradient boosting machine, a deep neural network, or even a proprietary API endpoint without modification. This agnostic nature makes it a universal tool for auditing and comparing explanations across entirely different model families within a single enterprise workflow.

The raw features used by a complex model are often uninterpretable to a human, such as word embeddings or pixel values. LIME bridges this gap by requiring the user to define an interpretable representation for the input. For text classification, this might be the presence or absence of individual words (a bag-of-words). For image classification, it might be the presence or absence of contiguous super-pixels. The explanation is then presented in this human-friendly space, such as highlighting the words that most contributed to a document being classified as fraudulent or the super-pixels that led to a specific image recognition.

Human comprehension degrades rapidly with complexity. LIME enforces explanation sparsity by explicitly limiting the number of features the local surrogate model can use, typically via L1 regularization (Lasso). The output is a concise list of the top-k features with their positive or negative contributions to the prediction. For a fraud analyst reviewing a blocked transaction, this means receiving a clear, actionable reason code like:

• Transaction Amount: +0.45 (increased fraud score)
• Time Since Last Login: -0.30 (decreased fraud score)
• Device Fingerprint Mismatch: +0.52 (increased fraud score) This sparsity constraint directly addresses the need for rapid, auditable decision justification.

While LIME is a local method, the SP-LIME (Submodular Pick LIME) algorithm extends it to provide a form of global understanding. It addresses the question: 'If you can only inspect a few instances, which ones should you look at?' SP-LIME uses a submodular optimization function to select a diverse, representative set of individual instances and their local explanations. This non-redundant set captures the main components of the model's global behavior, allowing a model risk manager to efficiently audit a model's logic across different regions of the feature space without examining every single prediction.

A key practical consideration for LIME is the stability-fidelity trade-off. Because the local neighborhood is generated through random perturbation, running LIME twice on the same instance can produce slightly different explanations. The fidelity of the local surrogate to the black-box model is measured by the R-squared value of the interpretable model on the perturbed samples. A high fidelity score indicates the surrogate accurately mimics the black-box in that locality. Increasing the number of perturbations improves stability and fidelity but increases computational cost, a critical factor when generating real-time explanations in a high-throughput fraud scoring pipeline.