Enhancing Visual Feature Attribution via Weighted Integrated Gradients

A new technique called Weighted Integrated Gradients (WG) aims to improve how computer vision models highlight the features that influence their decisions, addressing a known weakness in existing explanation methods [1]. Integrated Gradients (IG) is a standard tool in explainable AI for attributing a model's prediction to its input features, but its reliability depends heavily on the choice of a reference, or baseline, image [1]. A common multi-baseline extension, Expected Gradients (EG), attempts to reduce this sensitivity by averaging attributions across multiple baselines, yet it treats every baseline as equally informative [1]. In high-dimensional vision models, this uniform weighting can produce noisy or unstable explanations [1]. The new WG method, detailed in a paper by Tran Duc T, Tam Nguyen Trong, and Anh Nguyen Duc, introduces an unsupervised criterion to evaluate baseline suitability without requiring any annotation or domain knowledge [2][4]. This fitness function assesses each baseline solely from model behavior, and an efficient algorithm computes the score in logarithmic time relative to the number of baselines, enabling the approach to scale to large baseline sets [3]. The scores are then normalized into weights used to aggregate IG attributions, giving more influence to baselines that are more informative for a specific input [3]. The method preserves the core axiomatic properties of IG, including completeness, sensitivity, and implementation invariance, while providing a probabilistic justification for its weighting scheme under a fitness-relevance monotonicity assumption [1][4]. Experiments on standard vision benchmarks and common architectures, including convolutional networks and Transformers, showed that WG consistently outperformed EG, yielding improvements in attribution fidelity of up to 36 percent [1][4]. The authors note that these gains come with an additional computational cost for fitness evaluation, positioning WG as an attribution-fidelity trade-off rather than a faster alternative to EG [1][4]. The weighting mechanism also identifies informative baseline subsets, which reduces unnecessary variability and can lower computational costs by filtering out less useful baselines while maintaining high attribution accuracy [2][3].