Exponential Kernel

The kernel assigns a weight to each perturbed sample based on its proximity to the original instance. Samples that are closer receive weights approaching 1, while those farther away decay exponentially toward 0. This enforces the locality constraint, ensuring the surrogate model learns the decision boundary only in the immediate neighborhood rather than globally.

The weight is computed as: w = exp(-D(x, x')² / σ²)

• D(x, x'): A distance function between the original instance and the perturbed sample
• σ (sigma): The kernel width hyperparameter controlling the neighborhood size
• The squared distance in the numerator ensures rapid decay, creating a sharply localized focus
• For text, cosine distance on TF-IDF vectors is commonly used; for images, distance over superpixel presence

The kernel width is the critical tuning parameter that determines the effective radius of the local neighborhood:

• Small σ: Creates a tight, highly local neighborhood. High fidelity near the instance but unstable explanations due to limited samples
• Large σ: Creates a broader, smoother neighborhood. More stable but may capture non-local behavior, violating the locality principle
• OptiLIME frameworks automate this selection by balancing fidelity and stability across multiple runs

The distance function D(x, x') must be chosen to match the data type:

• Tabular Data: Euclidean distance on standardized numerical features, with Gower's distance for mixed categorical and continuous variables
• Text Data: Cosine distance on TF-IDF or embedding vectors, measuring semantic similarity independent of document length
• Image Data: Hamming-like distance over binary superpixel presence vectors, where each superpixel is either present or masked

The exponential kernel transforms standard linear regression into locally weighted regression. Each perturbed sample's contribution to the loss function is multiplied by its kernel weight. This forces the surrogate model—typically a sparse linear model using Lasso regression—to prioritize fitting points near the instance while ignoring distant, irrelevant perturbations that may cross decision boundaries.

The kernel directly mediates the fidelity-interpretability trade-off:

• A sharply decaying kernel produces high local fidelity but may overfit to noise in the immediate neighborhood
• A wider kernel produces smoother, more stable explanations but risks including samples from different decision regions
• The exponential form is preferred over linear decay because it creates a principled, smooth transition zone rather than a hard cutoff

A critical hyperparameter that controls the effective size of the local neighborhood. It determines how quickly sample weights decay as a function of distance from the original instance.

• A small kernel width creates a tight, highly local neighborhood, capturing fine detail but risking instability.
• A large kernel width smooths the explanation over a broader region, increasing stability but potentially missing sharp local variations.
• Selecting the optimal width balances the fidelity-interpretability trade-off.

The standard proximity measure used in LIME for text data. It calculates the similarity between two documents based on the angle between their TF-IDF vector representations.

• Cosine distance is magnitude-invariant, meaning it ignores differences in document length and focuses purely on the relative distribution of terms.
• A distance of 0 indicates identical directional vectors, while a distance of 1 indicates orthogonal, completely dissimilar content.
• This metric is essential for weighting perturbed text samples where random word removal changes vector magnitude.

The default proximity measure for tabular and image data in LIME. It calculates the straight-line distance between the original instance and a perturbed sample in the feature space.

• For tabular data, it operates on the standardized numerical feature vectors.
• For images, it measures the distance between the original and a masked version in the superpixel-presence space.
• The exponential kernel transforms this raw distance into a weight between 0 and 1, where a distance of 0 yields a weight of 1.

The statistical backbone of the LIME algorithm. This non-parametric method fits a simple surrogate model to a localized subset of data by solving a weighted least-squares problem.

• Each perturbed sample is assigned a weight by the exponential kernel based on its proximity to the target instance.
• The regression objective minimizes the weighted sum of squared errors, forcing the surrogate to prioritize fitting samples that are closer to the original instance.
• This ensures the explanation is a local approximation of the complex decision boundary, not a global one.

An optimization framework that automatically selects the optimal kernel width for a LIME explanation. It eliminates the need for manual hyperparameter tuning.

• It frames the selection as a constrained optimization problem that balances two competing objectives: maximizing local fidelity and maximizing explanation stability.
• Stability is measured by the concordance of feature importance rankings across multiple runs with different random seeds.
• OptiLIME provides a principled, data-driven alternative to ad-hoc kernel width selection, ensuring robust and reproducible explanations.

The property that a local explanation should remain consistent across multiple runs with different random perturbation samples. It is a direct measure of the explanation's reliability.

• An unstable explanation, where important features change drastically between runs, indicates the kernel width is too small or the number of samples is insufficient.
• The exponential kernel's width is the primary control for stability; a wider kernel averages over more samples, reducing variance.
• Stability is a prerequisite for trust; an explanation that changes randomly cannot be used for rigorous debugging or compliance auditing.