The Bias of Baseline Selection in Feature Attribution

328 concepts168 papers1 proposal

To understand why a deep learning model made a specific prediction, we often ask a counterfactual question: What would happen if this feature were missing? If removing a pixel, a word, or a medical biomarker changes the model's output significantly, we attribute high importance to that feature.

As formalized by Mukund Sundararajan et al. in Axiomatic Attribution for Deep Networks, feature attribution is fundamentally a comparative task. To quantify the impact of an input, we must measure it against a neutral reference point—a baseline—that represents the complete absence of information. Similarly, Avanti Shrikumar demonstrated in Learning Important Features Through Propagating Activation Differences that comparing neuron activations against a predefined reference state allows attribution methods to calculate meaningful contribution scores even when gradients are completely saturated.

However, defining what "absence" means in a deep network is surprisingly difficult. Whether we choose to replace features with zeros, solid colors, mean values, or blurred pixels, our choice of baseline introduces severe biases that can completely alter or invalidate the resulting explanation.

The Illusion of Neutrality

The core issue is that neural networks do not possess an innate concept of "nothingness." Every input configuration, including a blank image or a zeroed-out embedding vector, is processed as an active set of features.

As Lukas Brunke et al. observed in Evaluating Input Perturbation Methods for Interpreting CNNs and Saliency Map Comparison, neural networks treat solid backgrounds or blurred regions as distinct input features rather than a true absence of information. Because the model's internal logic reacts to these artificial backgrounds, the resulting saliency maps become highly sensitive to the specific baseline chosen for masking.

This sensitivity creates a profound reliability problem. Bob L. Sturm demonstrated this empirically in Reliable Local Explanations for Machine Listening by evaluating local explanation methods on audio models. Switching the masking strategy from zero-padding to mean-value imputation yielded entirely different sets of influential features. When an explanation depends more on the arbitrary choice of a baseline than on the model's internal logic, it introduces a severe risk of unintentional cherry-picking.

The Mathematical and Distributional Traps

When we force a fixed baseline into an attribution method, we typically trigger three distinct failure modes: blind spots, out-of-distribution artifacts, and systematic bias.

1. Blind Spots

Path-based attribution methods measure the accumulation of gradients along a trajectory from the baseline to the input. However, if an important feature in the input happens to share the exact same value as the baseline, its mathematical difference is zero. As Michael Terry pointed out in XRAI: Better Attributions Through Regions, these methods suffer from severe blind spots when input features match the reference baseline (such as a dark object against an all-black background reference), failing to assign any importance to critical features.

2. Out-of-Distribution (OOD) Artifacts

Replacing parts of an image or text sequence with arbitrary constants creates hybrid inputs that look nothing like the training data. Gary S. W. Goh argued in Understanding Integrated Gradients with SmoothTaylor for Deep Neural Network Attribution that fixed baselines are frequently statistical outliers relative to the model's training distribution. When a model evaluates these OOD inputs, its internal activations behave erratically. Jaap Jumelet et al. confirmed this behavior in language models in Feature Interactions Reveal Linguistic Structure in Language Models, showing that using zero-vectors or padding tokens to simulate feature removal creates out-of-distribution inputs that yield unfaithful and unstable interaction explanations.

3. Systematic Bias

Evaluating feature importance by replacing inputs with a single fixed value inherently penalizes features that naturally sit close to that value. In Evaluations and Methods for Explanation through Robustness Analysis, Cheng-Yu Hsieh et al. proved that traditional feature-importance metrics are systematically biased against features resembling the baseline. When a feature's original value is close to the reference value, its contribution is incorrectly minimized or ignored entirely, regardless of its actual role in the model's decision.

Information Leakage through the Mask

A more insidious problem arises in instance-wise feature selection, where an auxiliary "selector" model learns a binary mask to isolate a minimal sufficient subset of features for a "predictor" model. Instead of forcing the predictor to rely on the unmasked features, the selector often learns to cheat by encoding information directly into the mask's geometry.

Andrey Zhmoginov et al. discovered this phenomenon in Information-Bottleneck Approach to Salient Region Discovery. Saliency generators can inadvertently encode the target class label directly into the geometric shape or silhouette of the mask itself. The downstream classifier then bypasses the intended information bottleneck entirely, reading the mask's boundary rather than the underlying image content.

[Input Image] ──> [Selector Model] ──> [Mask Geometry (Leaks Label)]
                                                 │
                                                 ▼
[Masked Image] ──────────────────────> [Predictor Model] ──> [High Accuracy]
                                       (Reads mask shape,
                                        ignores pixels)

Harrie Oosterhuis et al. formalized this vulnerability in Local Feature Selection without Label or Feature Leakage for Interpretable Machine Learning Predictions. They demonstrated that mathematically impossible performance levels in selector-predictor architectures are often driven by this hidden communication channel. The selector encodes labels or hidden features into the specific configuration of the chosen mask, meaning the predictor identifies the answer based on which features were chosen rather than the actual values of those features.

Designing Better Reference States

To resolve baseline sensitivity, researchers have shifted away from fixed, arbitrary constants toward dynamic, distribution-aware reference states.

Path and Distribution Averaging

Instead of integrating gradients from a single, potentially out-of-distribution baseline, modern approaches smooth out the dependency. Fadi Thabtah et al. proposed Path-sampled integrated gradients, a framework that stochastically averages over multiple baselines along a path. This approach eliminates the arbitrary nature of a single reference point and significantly reduces attribution noise.

Generative Counterfactual Baselines

Rather than using uninformative black or blurred images that yield generic "one-vs-all" explanations, recent techniques generate context-aware baselines. Sheng-Min Shih et al. introduced GANMEX: One-vs-One Attributions using GAN-based Model Explainability, which leverages Generative Adversarial Networks to transform an input image into a realistic counterfactual baseline belonging to a specific target class. By comparing an input against a realistic alternative within the data distribution, the explanation highlights the exact discriminative features that distinguish two categories.

Symmetry-Aware Baselines

Baselines must also respect the mathematical properties of the model being explained. Jonathan Crabbé et al. established in Evaluating the Robustness of Interpretability Methods through Explanation Invariance and Equivariance that if a baseline input does not respect the model's underlying symmetry groups, the resulting explanations will fail to transform predictably when the input data is rotated, translated, or otherwise transformed. Ensuring baseline equivariance is now recognized as a requirement for robust, geometric interpretability.

Key Foundational Work

The (Un)reliability of Saliency Methods (2017) — contributed 6 concept elements to this direction.
Axiomatic Attribution for Deep Networks (2017) — contributed 3 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 193 papers
  2. Version 1 · 2026-10-04 · 235 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_baseline_sensitivity_ffaeb081,
  title = {baseline sensitivity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/baseline-sensitivity/rdr_ffaeb0814f4546b89481c01f74c67de7},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_baseline_sensitivity_ffaeb081_v2,
  title = {baseline sensitivity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/baseline-sensitivity/rdr_ffaeb0814f4546b89481c01f74c67de7/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake