The effect of whitening on explanation performance

Whitening can help, but inconsistently: While data whitening (a preprocessing technique that removes correlations between features) can improve the reliability of AI explanations in some cases, the improvement varies significantly depending on which explanation method and model type are used—meaning there's no one-size-fits-all solution.

Current explanation methods have a fundamental flaw: Popular feature attribution methods often incorrectly identify non-informative variables as important due to "statistical suppression" caused by correlated features, which can lead to serious misinterpretations of how AI models actually make decisions.

Preprocessing matters more than previously recognized: The study demonstrates that how data is prepared before training an AI model significantly affects whether explanations are trustworthy, suggesting that responsible AI development requires careful attention to preprocessing choices, not just the explanation method itself.

Synthetic benchmarks are essential for validation: The research used the XAI-TRIS benchmark with ground-truth data to rigorously test explanations, highlighting that organizations deploying AI should demand similar rigorous testing of explanation methods rather than assuming popular techniques are reliable.

No simple fix exists yet: The complex relationship between data structure, model architecture, and explanation accuracy means that practitioners cannot simply apply whitening and trust their explanations are now correct—ongoing validation and method-specific evaluation remain necessary.