New Framework Speeds Up Anchors Explanation Technique by Reusing Prior Rules

Anchors, a widely-used technique for generating local explanations of machine learning model predictions, has faced a persistent challenge: it's too slow for real-world deployment. A new research paper from arXiv addresses this bottleneck with a technique called MAnchors, which accelerates the explanation process by storing and repurposing rules generated for previous inputs.

The Problem with Anchors#

The Anchors algorithm works by iteratively refining explanations—sets of features that justify a model's decision for a specific input—until they reach the desired level of precision. While this approach is both model-agnostic and interpretable, the repeated computation across different inputs makes it computationally expensive, limiting its use in time-sensitive applications where users expect explanations on demand.

How MAnchors Works#

The proposed framework operates by maintaining a memory bank of low-precision, high-coverage rules learned from past explanations. Rather than starting from scratch for each new input, MAnchors adapts these cached rules through two complementary transformations:

• Horizontal transformation: Modifies a previously-generated explanation by swapping out features to match the characteristics of a new input
• Vertical transformation: Takes these adapted rules and refines them further until they achieve the precision required for the current prediction

This two-level approach lets the system leverage computational work already completed, dramatically reducing the time needed to generate new explanations.

Evaluation and Impact#

The researchers tested MAnchors across three distinct domains—tabular data, text, and images—and found that the method consistently reduced explanation generation time while preserving both the fidelity and interpretability that make Anchors valuable in the first place. By closing the efficiency gap, this work aims to make model-agnostic explanations practical for applications where speed matters, such as customer-facing systems or regulatory audits requiring rapid justifications of automated decisions.