Regional Explanations: Bridging Local and Global Variable Importance

Researchers identify fundamental flaws in Local Shapley Values and LIME, two widely-used machine learning interpretation methods that fail to reliably detect locally important features. They propose R-LOCO, a new approach that bridges local and global explanations by segmenting input space into regions and applying global attribution methods within those regions for more faithful local attributions.

This research addresses a critical gap in machine learning interpretability, a field increasingly vital as AI systems make high-stakes decisions in finance, healthcare, and other regulated industries. The authors demonstrate that two of the most popular local attribution methods—Local Shapley Values and LIME—systematically fail to correctly identify which features actually influence individual predictions, even under ideal conditions. This is more than an academic concern; flawed explanations can lead practitioners to misunderstand model behavior and make poor deployment decisions.

The core issue stems from how these methods handle feature independence and statistical relationships. Both methods can assign importance to features that have no causal effect on predictions or that only correlate with truly relevant features through statistical association. This violates basic principles of sound attribution: a feature should only matter if it directly influences output or depends on features that do.

R-LOCO's innovation lies in its regional segmentation approach, which treats the input space as composed of zones with consistent feature importance patterns. By applying global methods within these regions rather than purely locally, it captures instance-specific details while maintaining the stability and fidelity that global methods provide. This hybrid approach offers practical benefits for practitioners: more trustworthy explanations that align with actual model behavior.

For the AI industry, particularly in regulated sectors requiring model transparency, this work signals that existing explanation tools may be less reliable than assumed. Organizations using LIME or Shapley values for compliance or decision-making should consider validation studies. The research pushes the field toward more principled attribution methods, likely influencing future regulatory requirements around AI explainability standards.

• →Local Shapley Values and LIME assign importance to features that don't actually influence predictions, violating fundamental attribution principles
• →R-LOCO bridges local and global explanations by segmenting input space into regions with similar feature importance characteristics
• →The proposed method delivers more faithful attributions while avoiding instability issues common in purely local explanation approaches
• →This work has significant implications for regulated industries relying on current interpretability methods for compliance and accountability
• →Organizations using LIME or Shapley values should validate their explanations against actual model behavior in high-stakes applications