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

TL;DR

This work tackles false-positive attributions caused by suppressor variables in explanations of non-linear models. It introduces PatternLocal, which converts local discriminative surrogates into a generative pattern by kernel-weighted regression, suppressing suppressors while preserving local fidelity. The method is evaluated on XAI-TRIS, an artificial MRI lesion benchmark, and EEG motor imagery, where it consistently reduces suppressor-driven attributions and yields physiologically plausible insights. PatternLocal thus enhances the reliability and actionability of local explanations without retraining models, with strong implications for high-stakes domains.

Abstract

Suppressor variables can influence model predictions without being dependent on the target outcome, and they pose a significant challenge for Explainable AI (XAI) methods. These variables may cause false-positive feature attributions, undermining the utility of explanations. Although effective remedies exist for linear models, their extension to non-linear models and instance-based explanations has remained limited. We introduce PatternLocal, a novel XAI technique that addresses this gap. PatternLocal begins with a locally linear surrogate, e.g., LIME, KernelSHAP, or gradient-based methods, and transforms the resulting discriminative model weights into a generative representation, thereby suppressing the influence of suppressor variables while preserving local fidelity. In extensive hyperparameter optimization on the XAI-TRIS benchmark, PatternLocal consistently outperformed other XAI methods and reduced false-positive attributions when explaining non-linear tasks, thereby enabling more reliable and actionable insights. We further evaluate PatternLocal on an EEG motor imagery dataset, demonstrating physiologically plausible explanations.

Figures (19)

• Figure 1: Conceptual step-by-step overview of how the PatternLocal method generates local explanations. A classifier is trained on the dataset (a), with both the dataset and the model’s decision boundary visualized in a 2D projection. Many XAI methods produce importance maps (c) through local linearization (e.g., LIME, KernelSHAP). PatternLocal enhances this process by transforming the local discriminative surrogate into a generative explanation (d), thereby reducing the influence of irrelevant or misleading features. The resulting importance maps (c–d) can be directly compared with the ground-truth attributions (b).
• Figure 2: Feature–importance comparison on the XOR toy problem. We draw $2\,500$ i.i.d. samples from the generative process of \ref{['eq:toy_data_generation']}, so the label depends solely on the interaction between $x_1$ and $x_2$, while $x_3$ is a suppressor that carries no marginal predictive signal. Each plot shows $(x_1,x_2)$ pairs, colored by the local normalized feature importance that four XAI methods assign to the suppressor variable $x_3$ when applied to the smooth classifier. Every plot reports the empirical mean magnitude of this attribution across all test points. Whereas LIME, KernelSHAP, and Gradient all attribute substantial importance to $x_3$, PatternLocal correctly drives the attribution of $x_3$ to (almost) zero.
• Figure 3: Examples of $64 \times 64$ instances from the XAI-TRIS Benchmark across six different scenario types. Each row represents a distinct structural pattern (LIN, XOR, RIGID), and each column pair shows a binary label class ($y = 0$ and $y = 1$) along with the corresponding instance and attribution mask. On the right (\ref{['fig:example-data-a']}), correlated noise introduces a spurious suppressor effect, while on the left (\ref{['fig:example-data-b']}), white noise does not.
• Figure 4: Earth mover's distance (EMD)
• Figure 5: Importance mass error (IME)