LIME

LIME operates as a black-box explainer that treats the underlying model as a complete abstraction. It requires no access to internal weights, gradients, or architecture details—only the ability to query the model with inputs and receive predictions. This makes it applicable to any classifier or regressor, including deep neural networks, gradient-boosted trees, support vector machines, and even proprietary APIs. The technique works by generating a local neighborhood of perturbed samples around the instance to be explained, weighting them by proximity, and training a simple interpretable surrogate model on this local dataset.

LIME explicitly prioritizes local fidelity over global understanding. Rather than attempting to explain the entire model's behavior across all input space—an intractable problem for complex models—it focuses on producing an explanation that is faithful to how the model behaves in the immediate vicinity of a single prediction. The objective function balances two competing goals:

• Interpretability: The surrogate model must be simple enough for a human to understand (e.g., a sparse linear model or shallow decision tree)
• Local Fidelity: The surrogate must accurately mimic the complex model's decision boundary in the local region This trade-off is formalized in the LIME optimization framework.

The core mechanism of LIME involves creating a synthetic local neighborhood through controlled perturbation of the original instance. For structured data like text or tabular features, perturbations include:

• Text: Randomly removing words or tokens from the input document
• Tabular: Sampling values from a normal distribution centered on the original feature values
• Images: Segmenting the image into superpixels and randomly turning contiguous regions on or off Each perturbed sample is weighted by an exponential kernel based on its distance from the original instance, ensuring that samples closer to the point of interest have greater influence on the surrogate model's training.

LIME introduces a critical distinction between the model's native feature space and an interpretable representation. For text classification, the interpretable representation is typically a binary bag-of-words indicating the presence or absence of individual words. For image classification, it is the presence or absence of contiguous superpixel segments. The surrogate model is trained on this binary interpretable space, not the original pixel or embedding space. This mapping ensures that the final explanation—a weighted list of the most influential interpretable components—is directly meaningful to a human operator, showing exactly which words or image regions drove the prediction.

The default interpretable surrogate model in LIME is a sparse linear model trained with Lasso (L1) regularization. Sparsity is enforced by selecting only the top K features to include in the explanation, where K is a user-defined parameter controlling explanation complexity. The resulting explanation takes the form:

• A list of the most influential interpretable components
• Each component assigned a positive or negative weight indicating its contribution toward or against the predicted class
• The sum of all weights plus the intercept approximates the complex model's prediction probability for that instance This format is immediately intelligible: 'Prediction X because of features A, B, and C.'

While LIME is fundamentally a local explanation method, it includes a companion algorithm called Submodular Pick (SP-LIME) for selecting a representative set of instances to explain globally. SP-LIME frames instance selection as a submodular optimization problem with a budget constraint, maximizing the coverage of important features across the chosen explanations. The goal is to provide a non-redundant, diverse set of individual explanations that collectively offer insight into the model's overall behavior. This addresses the limitation that explaining random individual instances may miss critical global patterns in the model's decision logic.

An attribution method that satisfies the sensitivity and implementation invariance axioms that LIME's local approximations may violate. It computes feature importance by accumulating gradients along a straight-line path from a baseline input (typically zeros or noise) to the actual input. For RF signals, this baseline might be a silent channel or thermal noise floor, revealing which signal components drive classification.

While LIME explains why a prediction was made, counterfactuals explain how to change it. This method identifies the minimal perturbation to an input required to flip the model's decision to a desired alternative. In spectrum sensing, a counterfactual might reveal: 'Increase the signal-to-noise ratio by 2 dB at 2.4 GHz to reclassify this emitter as friendly.'

A decomposition-based technique that redistributes a neural network's prediction score backwards through layers using a conservation property—total relevance is preserved from output to input. Unlike LIME's model-agnostic sampling, LRP is specific to deep networks but provides finer-grained attribution. For RF fingerprinting, LRP can highlight which transient samples in a burst most influenced identification.

A gradient-based visualization technique originally designed for CNNs that produces coarse localization maps. When applied to RF spectrograms or time-frequency representations, Grad-CAM highlights which regions of the spectrum most influenced a modulation classification decision. This complements LIME by providing spatial context that per-sample perturbation analysis may miss.

LIME explanations can expose vulnerabilities in RF models by revealing reliance on spurious correlations. Adversarial robustness is the property of maintaining correct predictions under intentionally perturbed inputs—critical when an adversary transmits waveforms designed to fool your classifier. Understanding LIME's local decision boundaries helps engineers harden models against evasion attacks over the air.