Explanation Stability

The perturbation sampling process in LIME relies on random number generation to create synthetic data points. Different seeds produce different neighborhoods, which can lead to divergent feature importance rankings for the same prediction.

• A stable explanation should maintain consistent top features across seeds
• High variance in feature weights signals unreliable explanations
• Seed sensitivity is a primary diagnostic for explanation trustworthiness

The kernel width hyperparameter controls the effective size of the local neighborhood by determining how quickly sample weights decay with distance from the original instance.

• Too narrow: surrogate model overfits to very few samples, causing high variance
• Too wide: explanation loses local fidelity and captures global trends
• OptiLIME automates kernel width selection by balancing fidelity and stability

The number of perturbed samples generated around the instance directly impacts explanation stability. Insufficient samples lead to high-variance surrogate models that produce inconsistent feature attributions.

• More samples reduce variance but increase computational cost
• Stability plateaus after a data-dependent threshold
• Text explanations typically require more samples than tabular data due to higher dimensionality

The L1 regularization (Lasso) penalty in the sparse linear surrogate model enforces sparsity but can amplify instability. Small changes in the perturbed dataset can cause different features to be selected.

• Stronger regularization increases sparsity but may reduce stability
• Bayesian LIME addresses this by providing uncertainty estimates alongside feature weights
• Explanation regularization techniques can explicitly penalize variance across runs

Stability characteristics vary significantly by data type due to differences in perturbation strategies and interpretable representations.

• Text data: token masking creates discrete, high-dimensional neighborhoods; TF-IDF cosine distance weighting affects stability
• Image data: superpixel segmentation randomness introduces additional variance beyond perturbation sampling
• Tabular data: continuous feature discretization into quantiles can stabilize explanations by reducing the feature space

Quantitative metrics exist to measure explanation stability across multiple runs with different random seeds.

• Jaccard index measures overlap between top-k feature sets across runs
• Rank correlation (Spearman's ρ) assesses consistency of feature importance ordering
• Feature weight variance directly quantifies the dispersion of individual feature attributions
• Running 10-50 repetitions and visualizing weight distributions is standard practice for auditing stability

A critical hyperparameter controlling the effective size of the local neighborhood by determining how quickly sample weights decay with distance from the original instance.

• A narrow kernel focuses on very close perturbations, increasing local fidelity but often reducing stability due to high variance in the small sample.
• A wide kernel smooths the explanation over a larger region, improving stability but potentially missing sharp local decision boundaries.
• Selecting the optimal kernel width is the primary mechanism for balancing the fidelity-stability trade-off.

An optimization framework that automatically selects the optimal kernel width for a LIME explanation.

• It operates by balancing the trade-off between local fidelity and the stability of the explanation across different random seeds.
• The algorithm searches for the largest kernel width that still maintains a predefined fidelity threshold, maximizing stability without sacrificing accuracy.
• This removes the need for manual hyperparameter tuning, providing a more objective and reproducible explanation pipeline.

The process of generating a synthetic neighborhood of data points by randomly altering or masking features of the original instance.

• Explanation instability often originates here: different random seeds produce different sets of perturbed samples.
• Sampling strategies vary by modality—token masking for text, superpixel masking for images, or Gaussian perturbation for tabular data.
• Increasing the number of samples generally improves stability by providing a more comprehensive view of the local decision boundary, at the cost of higher computation.

An extension of LIME that uses Bayesian ridge regression as the surrogate model instead of standard linear regression.

• It provides uncertainty estimates alongside feature importance scores, indicating the confidence of the explanation itself.
• High variance in the posterior distribution of feature weights signals explanation instability.
• This allows practitioners to distinguish between robust, high-confidence feature attributions and those that are artifacts of the random sampling process.

An interpretable surrogate model that uses Lasso (L1) regularization to select only a small number of the most important features.

• The sparsity constraint directly impacts stability: a model that selects different features across runs exhibits feature selection instability.
• Ridge regression (L2) produces denser but often more stable explanations by distributing weight smoothly across correlated features.
• The choice of regularization is a key architectural decision in the explanation pipeline, trading conciseness for consistency.

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

• It ignores differences in document length, focusing purely on the semantic similarity of word distributions.
• The distance metric directly determines which perturbed samples are weighted most heavily in the local regression.
• An inappropriate distance metric can introduce systematic bias into the neighborhood definition, undermining the stability and validity of the resulting explanation.