Explainable Artificial Intelligence for Dependent Features: Additive Effects of Collinearity

Current XAI methods may give misleading explanations when features are correlated: Most existing explainable AI approaches assume features work independently, but in real-world data, features are often correlated with each other. This can lead to inaccurate explanations of how AI models make decisions.

Collinearity is a widespread but underaddressed problem in AI transparency: The paper identifies collinearity (when input features are statistically dependent) as "one of the big issues that XAI methods face," suggesting this is a common challenge that hasn't received adequate attention in the responsible AI community.

A new method (AEC) offers more reliable explanations for correlated data: The proposed Additive Effects of Collinearity approach breaks down complex models into simpler components to better understand how correlated features jointly influence predictions, showing improved robustness compared to existing methods in both simulated and real-world tests.

Better feature explanations matter for trustworthy AI deployment: When organizations use XAI methods to understand their AI systems, flawed explanations due to collinearity could lead to incorrect conclusions about which factors drive decisions—potentially affecting fairness audits, regulatory compliance, and stakeholder trust in AI systems.