A new academic method wants to patch a quiet flaw in how we explain AI decisions: duplicate or overlapping data.
Researchers propose MCIR-M (Mutual Correlation Impact Ratio Method), a dependence-aware approach to measuring which features actually matter in a model's predictions. It targets a known weak spot in popular tools like SHAP, LIME, HSIC, MI/CMI, and SAGE, which can produce unstable rankings when features are highly correlated or nearly identical. MCIR works by conditioning each feature on its most dependent neighbors, then computing a ratio of conditional to block-level information - a score from 0 to 1 that drops to zero when a feature is fully redundant. The team also built a lighter version that estimates these scores from a fraction of the data, and tested the approach on synthetic redundancy experiments and the UCI Human Activity Recognition benchmark.
This matters because real-world data is rarely clean. Sensor readings, financial indicators, and medical records are full of overlapping signals, and when explainability tools can't agree on which overlapping feature "mattered," the resulting rankings become unreliable right when people lean on them most, like auditing a model's decisions.
The catch: against real-data baselines, MCIR's results were mixed, not a clean win. Call it a useful new diagnostic for a specific failure mode, not a replacement for SHAP.