Minimizing False-Positive Attributions in Explanations of Non-Linear Models

Suppressor variables are a hidden problem in AI explanations: The paper identifies that certain variables can influence what a model predicts without actually being related to the outcome being predicted. These "suppressor variables" trick popular explanation methods into falsely identifying them as important—a problem that has been largely overlooked in non-linear AI systems used in real-world applications.

Current explanation methods have significant blind spots: Existing techniques like LIME and SHAP work reasonably well for simple linear models but struggle with complex, non-linear models (like deep learning systems). This gap means explanations for many modern AI systems may be unreliable without users realizing it.

PatternLocal offers a practical solution: The researchers developed a new technique that takes existing explanation methods and improves them by converting their outputs into a different mathematical form that filters out suppressor variables. Testing showed it consistently reduced false-positive attributions compared to other methods.

Improvements were validated rigorously: The method was tested on a standardized XAI benchmark and also on real-world EEG brain imaging data, where the explanations aligned with known neuroscience—suggesting the approach produces genuinely meaningful insights rather than just statistical artifacts.

Better explanations enable more trustworthy AI: By reducing false attributions, this work helps ensure that when organizations use AI explanations to make decisions (in healthcare, hiring, lending, etc.), they're actually understanding what the model is doing rather than being misled by spurious patterns.