Guidelines For The Choice Of The Baseline in XAI Attribution Methods

Cristian Morasso, Giorgio Dolci, Ilaria Boscolo Galazzo, Sergey M. Plis, Gloria MenegazarXiv 2025 · DOI · arXiv

Abstract

Explainable AI (XAI) methods, particularly Baseline Attribution Methods (BAMs) like Integrated Gradients, are highly sensitive to the choice of a "neutral" reference point. Arbitrary baseline selection—such as using a black image or Gaussian noise—can lead to misleading explanations where attribution signs lose physical meaning, a critical failure in high-stakes domains. We propose Informed Baseline Search (IBS), an architecture-agnostic algorithm that identifies an optimal baseline situated directly on the model's decision boundary. Our results demonstrate that baselines found via IBS provide more reliable, consistent, and interpretable feature attribution maps across both low- and high-dimensional datasets.

Introduction

In the field of XAI, BAMs explain a model's prediction by integrating gradients along a path from a reference baseline to the input sample. While mathematically sound, the practical application of these methods is often hindered by the arbitrary nature of baseline selection. If the path between the baseline and the sample crosses the model's decision boundary (DB) multiple times or traverses regions outside the training data manifold, the resulting attribution values can become "corrupted." This leads to inconsistent results where the sign of the attribution no longer reliably indicates whether a feature supports or opposes a specific classification.

Methods

The authors argue that a truly neutral baseline must reside on the model's decision boundary, where the class probability is exactly 0.5. To find such a point, they introduce Informed Baseline Search (IBS).

Conceptual diagram of the Informed Baseline Search (IBS) algorithm showing the iterative search for a boundary point.

IBS is an iterative algorithm that starts from an input sample and moves toward a randomly sampled point from the "losing" class. By adjusting the step size based on model confidence, it converges on the decision boundary. Furthermore, the authors define an Optimality Criterion: the ideal baseline is the orthogonal projection of the sample onto the DB. This minimizes the path distance and prevents the path from crossing non-linear boundary segments multiple times.

Key Insights: The "Crossing" Trap

A primary finding of this research is the danger of the "Crossing Trap." When a baseline is chosen far from the sample, the path may cross the decision boundary in complex ways. As shown in the gradient analysis below, paths to distant baselines can cause gradients to flip signs or cancel each other out, resulting in noisy attribution maps.

Comparison of data distributions and gradients along paths to far (purple) vs. close (green) baselines.

By ensuring the baseline is the closest point on the "inner" decision boundary (the part residing within the data manifold), the attributions maintain a consistent sign. This restores physical interpretability to the model's logic, allowing researchers to determine exactly how feature changes influence classification outcomes.

Key Results

  • Scalability and Robustness: IBS successfully identified decision boundaries in scenarios ranging from 2D synthetic clusters to high-dimensional neuroimaging simulations (5,290 features).
  • Architecture Agnostic: Unlike exact methods like SplineCAM, which are limited to specific architectures (e.g., no max-pooling), IBS is compatible with any differentiable model.
  • Attribution Quality: Experiments showed that optimal baselines produced strictly positive attributions for known informative features, whereas random baselines produced ambiguous, mixed-sign results.
  • Efficiency: The GPU-friendly implementation of IBS offers a significantly lower computational footprint compared to existing boundary-mapping techniques.

Conclusion and Future Work

Informed Baseline Search provides a principled framework for baseline selection in XAI. By anchoring explanations to the model's decision boundary, IBS ensures that feature attributions are both mathematically consistent and physically meaningful. Future work will focus on applying IBS to complex real-world biomedical datasets and exploring its utility in generating counterfactual explanations.

Summary updated

BibTeX
@misc{morasso2025guidelines,
  author = {Cristian Morasso and Giorgio Dolci and Ilaria Boscolo Galazzo and Sergey M. Plis and Gloria Menegaz},
  title = {{Guidelines For The Choice Of The Baseline in XAI Attribution Methods}},
  year = {2025},
  eprint = {2503.19813},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2503.19813},
}
Figure