Anchors

Anchors generate if-then rules that achieve a user-specified precision threshold, typically 95% or higher. Unlike LIME's linear approximations, an anchor rule states that if specific conditions hold, the model's prediction is guaranteed to remain stable with high probability. This makes anchors ideal for regulatory compliance where explanations must be deterministic and verifiable.

• Precision guarantee: Configurable threshold ensures rule reliability
• Rule format: Human-readable conditions like 'If age > 30 AND income < $50k THEN predict default'
• Coverage metric: Measures how broadly the rule applies across instances

Anchors operate as a black-box explanation method, requiring only the model's prediction function and no access to internal weights or gradients. This makes it applicable to any classifier, including tree ensembles, deep neural networks, and proprietary APIs. The method perturbs the input instance by sampling from a perturbation distribution and queries the model to identify feature conditions that consistently anchor the prediction.

• Zero internal access: Works with any model exposing a predict function
• Framework-agnostic: Compatible with scikit-learn, TensorFlow, PyTorch, and cloud APIs
• Multi-class support: Explains predictions across all output classes

Anchors are built using a KL-LUCB bandit algorithm that efficiently searches for the shortest rule with sufficient precision. Starting from an empty rule, the algorithm iteratively adds candidate feature predicates, evaluating each using multi-armed bandit techniques to minimize model queries. This bottom-up approach ensures the final anchor is minimal in length while maintaining the precision guarantee.

• Bandit optimization: KL-LUCB algorithm minimizes computational cost
• Beam search candidate generation: Explores rule space efficiently
• Statistical stopping criteria: Terminates when precision bound is satisfied

The method generates synthetic neighborhood samples by perturbing the original instance according to a user-defined perturbation distribution. For tabular data, this may involve sampling from a multivariate Gaussian or using training data marginal distributions. For text, it replaces words with random alternatives or UNK tokens. The perturbation strategy directly impacts anchor quality and must reflect realistic data variations.

• Custom perturbation spaces: Define domain-appropriate sampling strategies
• Discrete and continuous support: Handles mixed feature types natively
• Instance-specific neighborhoods: Perturbations centered on the explained instance

Anchors explicitly balance precision (the fraction of perturbed instances satisfying the rule that share the prediction) against coverage (the fraction of all instances in the neighborhood that satisfy the rule). A high-precision anchor with low coverage explains only a narrow slice of instances, while broader coverage may sacrifice precision. This trade-off is user-configurable, allowing auditors to prioritize either strictness or generality.

• Precision: Proportion of correct predictions under the anchor rule
• Coverage: Proportion of the perturbation space covered by the anchor
• Configurable threshold: Set minimum acceptable precision before deployment

While LIME provides local linear approximations that may be unfaithful to the underlying model, anchors offer formal precision guarantees that LIME cannot. Unlike SHAP, which assigns continuous importance scores to all features, anchors produce discrete, actionable rules that are easier for non-technical stakeholders to interpret. Anchors complement these methods by providing a different explanatory lens focused on sufficiency conditions rather than feature attribution.

• vs LIME: Anchors provide verifiable guarantees; LIME provides weighted approximations
• vs SHAP: Anchors output rules; SHAP outputs feature importance scores
• Complementary use: Deploy anchors alongside attribution methods for comprehensive explanations