Research map / Learning theory and algorithms / Prediction and decision losses
Minimizing False-Positive Attributions in Explanations of Non-Linear Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces PatternLocal, which converts local linear surrogate weights (LIME, KernelSHAP, gradients) into kernel-weighted generative patterns via regression on the surrogate output, suppressing suppressor-driven false-positive attributions for non-linear models.
Problem
Post-hoc local explanations for non-linear models assign false-positive importance to suppressor variables that affect predictions without statistical dependence on the target; existing Pattern remedies only cover linear models.
Approach
Extends the linear Pattern (forward-model) transformation to local explanations: regress the simplified local inputs onto the local surrogate's output under a kernel weighting, yielding a covariance-based pattern that zeroes suppressor contributions in the toy XOR case while keeping local fidelity.
Key results
- On XAI-TRIS MLP with identity mapping, PatternLocal substantially outperforms other XAI methods in XOR and RIGID scenarios on EMD and IME.
- In RIGID CORR, PatternLocal beats most XAI methods but is comparable to or worse than the Sobel and Laplace filters.
- On EEG motor imagery, SSD dipole fit of PatternLocal explanations is 0.756 ± 0.090 versus 0.604 ± 0.013 for LIME.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PatternLocal instead of LIME | EEG Motor Imagery explanations · EEG Motor Imagery test set | dipole fit: 0.604 \pm 0.013 -> 0.756 \pm 0.090 | Helped |
| Use PatternLocal instead of other XAI methods | XAI-TRIS explanations · XAI-TRIS XOR and RIGID scenarios | EMD | Helped |
| Use PatternLocal instead of Sobel or Laplace | XAI-TRIS explanations · XAI-TRIS RIGID CORR | EMD and IME | Mixed |
| Use PatternLocal instead of LIME and filter-based methods | Artificial Lesion MRI explanations · Artificial Lesion MRI | EMD | Helped |
Surprising findings
- Simple edge-detection filters (Sobel, Laplace) match or beat PatternLocal on RIGID CORR and perform well on XAI-TRIS generally.
- On the EEG data, PatternLocal's dipole fit is similar to raw instances (0.738 ± 0.013).
Limitations
- Requires access to representative training data in the neighborhood of each explanation.
- Assumes spatial alignment/consistency across samples, which may fail in unconstrained natural images.
- Can still misattribute importance like other XAI methods.
Experimental setup
Datasets: XAI-TRIS Artificial Lesion MRI PhysioNet EEG Motor Movement/Imagery
Benchmarks: XAI-TRIS Artificial Lesion MRI PhysioNet EEG Motor Movement/Imagery
Metrics: Earth Mover's Distance Importance Mass Error dipole fit correlation
Baselines: LIME Integrated Gradients Saliency DeepLift GradientShap GuidedBackProp Sobel Laplace KernelSHAP
Builds on: Pattern (Haufe et al. 2014) LIME KernelSHAP PatternNet PatternAttribution
Closest papers by meaning
- CSBrain: A Cross-scale Spatiotemporal Brain Foundation Model for EEG Decoding (NeurIPS 2025 · similarity 0.84)
- A Local-Ascending-Global Learning Strategy for Brain-Computer Interface (AAAI 2024 · similarity 0.83)
- Inverse decision-making using neural amortized Bayesian actors (ICLR 2025 · similarity 0.83)
- BrainOmni: A Brain Foundation Model for Unified EEG and MEG Signals (NeurIPS 2025 · similarity 0.82)
- What do EEG Foundation Models Capture from Human Brain Signals? (NeurIPS 2026 · similarity 0.82)
- Cross-Modal Representational Knowledge Distillation for Enhanced Spike-informed LFP Modeling (NeurIPS 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.