Guidelines For The Choice Of The Baseline in XAI Attribution Methods

TL;DR

This paper tackles the fragility of attribution maps produced by baseline-guided XAI methods by analyzing how baseline choice affects explanations and proposing a practical solution. It introduces Informed Baseline Search (IBS), a DB-guided sampling algorithm that identifies the orthogonal projection of a sample onto the inner DB to serve as the optimal baseline for Integrated Gradients. Through synthetic experiments and comparisons with SplineCAM and DeepView, IBS demonstrates consistent DB localization and improved attribution faithfulness when the optimal BL is used, while highlighting the limitations of suboptimal baselines. The work provides clear guidelines for baseline selection, a GPU-friendly implementation, and a roadmap for extending the approach to broader BAMs and more realistic data, with significant implications for reliability and interpretability in biomedical AI applications.

Abstract

Given the broad adoption of artificial intelligence, it is essential to provide evidence that AI models are reliable, trustable, and fair. To this end, the emerging field of eXplainable AI develops techniques to probe such requirements, counterbalancing the hype pushing the pervasiveness of this technology. Among the many facets of this issue, this paper focuses on baseline attribution methods, aiming at deriving a feature attribution map at the network input relying on a "neutral" stimulus usually called "baseline". The choice of the baseline is crucial as it determines the explanation of the network behavior. In this framework, this paper has the twofold goal of shedding light on the implications of the choice of the baseline and providing a simple yet effective method for identifying the best baseline for the task. To achieve this, we propose a decision boundary sampling method, since the baseline, by definition, lies on the decision boundary, which naturally becomes the search domain. Experiments are performed on synthetic examples and validated relying on state-of-the-art methods. Despite being limited to the experimental scope, this contribution is relevant as it offers clear guidelines and a simple proxy for baseline selection, reducing ambiguity and enhancing deep models' reliability and trust.

Figures (9)

• Figure 1: Gradients behavior along the interpolated path between two different BLs, the purple one is randomly sampled on the DB, while the green BL is orthogonal to the sample. The red vertical line in the purple plot represents the location of the crossed DB. The lower axis of the line plots represents the X value, and the top axis represents the Y value.
• Figure 2: Two feature graphic example of the algorithm (three steps), the green star in the blue class (0) is the BL search starting point, and the red dotted line represents the DB. Running IBS on the starting point, as first, the losing class is 1, so it extracts the direction from BL to Class 1, then moves the BL over this direction for a given magnitude (green vector), in order to obtain the next BL candidate, then the prediction is in favor of class 1, so it does the same with class 0, so on and so forth until an optimal BL is reached.
• Figure 3: Dataset representation, starting from upper left Custom dataset, then Spiral dataset, and in the second row, Three features dataset and finally an image from the Simulated brain views.
• Figure 4: BL problem, with different BLs (First column, Feature 0 on X-axis and Feature 1 on Y-axis), the features attributions for the class 0 (Orange) are opposite (second column), then Delta (third column), and finally in the last column are reported the Cumulative Gradients (CG).
• Figure 5: Image case. Left image: important feature locations labeled as A, B, and C. First and second row: comparison between two BLs on IG attributions (similar to the previous figure). Bottom row: data distributions of the three highlighted features (A, B, C), BL location, and gradient directions. The green color is used for optimal BL and red for the random one.