Enhancing Visual Feature Attribution via Weighted Integrated Gradients

TL;DR

This work tackles the sensitivity of Integrated Gradients to baseline choice in computer vision by introducing Weighted Integrated Gradients (WG), which assigns input-dependent weights to baselines using an unsupervised fitness criterion. WG preserves IG’s core axioms while delivering higher attribution fidelity and stability than Expected Gradients, backed by a probabilistic analysis and finite-sample guarantees. An efficient $O(\log n)$ algorithm computes baseline fitness $D_\alpha$, enabling scalable weighting and optional filtering to reduce computation by substantial margins. Empirical results on ImageNet and COCO across multiple architectures show WG achieving meaningful improvements in attribution quality (e.g., 24.8% average Deletion AUC gain) and producing clearer saliency maps, making explanations more reliable for practical use in explainable AI for vision. The approach also demonstrates practical benefits through effective baseline selection and efficiency gains via filtering, broadening the applicability of gradient-based attribution in real-world models.

Abstract

Integrated Gradients (IG) is a widely used attribution method in explainable AI, particularly in computer vision applications where reliable feature attribution is essential. A key limitation of IG is its sensitivity to the choice of baseline (reference) images. Multi-baseline extensions such as Expected Gradients (EG) assume uniform weighting over baselines, implicitly treating baseline images as equally informative. In high-dimensional vision models, this assumption often leads to noisy or unstable explanations. This paper proposes Weighted Integrated Gradients (WG), a principled approach that evaluates and weights baselines to enhance attribution reliability. WG introduces an unsupervised criterion for baseline suitability, enabling adaptive selection and weighting of baselines on a per-input basis. The method not only preserves core axiomatic properties of IG but also provides improved theoretical guarantees on the quality of explanation over EG. Experiments on commonly used image datasets and models show that WG consistently outperforms EG, yielding 10 to 35 percent improvements in attribution fidelity. WG further identifies informative baseline subsets, reducing unnecessary variability while maintaining high attribution accuracy. By moving beyond the idea that all baselines matter equally, Weighted Integrated Gradients offers a clearer and more reliable way to explain computer-vision models, improving both understanding and practical usability in explainable AI.

Figures (2)

• Figure 1: The average AUC of Deletion Score for Filtered WG using ResNet50 architecture. Low-quality baselines were removed based on their fitness scores, resulting in more trustworthy final explanations, as reflected by a reduction in the AUC of the Deletion Score. The fitness score serves as an effective criterion to balance the trade-off between computational efficiency and explanation faithfulness
• Figure 2: Saliency map comparisons across different models (columns) and methods (rows). The top row is our proposed method, while the bottom row employs Expected Gradients. Our proposed method, Weighted Integrated Gradients (WG), consistently outperforms the baseline approach by focusing more sharply on the main object, demonstrating superior precision and relevance in feature attribution across all models.

Theorems & Definitions (4)

• Proposition 1: WG favors relevant baselines more than EG
• Theorem 2: High-probability relevance guarantee for WG
• proof : Proof sketch of Proposition \ref{['prop:wg_improves_relevance']}
• proof : Proof sketch of Theorem \ref{['thm:wg_probabilistic_rate']}