Enhancing Visual Feature Attribution via Weighted Integrated Gradients

2026-06-16 · Source: stat.ML updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

The study introduces Weighted Average Gradients (WG), a novel explainable AI technique that enhances visual feature attribution by addressing the limitations of traditional Integrated Gradients (IG) and Expected Gradients (EG). Unlike EG, which uniformly samples and averages baselines, WG unsupervisedly evaluates baseline suitability using a perturbation-based fitness function, D_alpha, and assigns proportional weights. This approach theoretically satisfies essential explanation criteria like completeness, sensitivity, and implementation invariance, while offering greater stability and variance reduction. Experimental results show WG outperforms EG by 10-35% on main metrics, including AUC of Deletion Score, and can filter effective baselines to reduce computational cost while maintaining high accuracy. The code is available at https://github.com/Tamnt240904/weighted_baseline.

Key takeaway

For AI Scientists and Machine Learning Engineers developing interpretable models, consider adopting Weighted Average Gradients (WG) to improve feature attribution accuracy and stability. Your current reliance on uniform baseline averaging in methods like Expected Gradients may introduce inconsistencies. Implementing WG, which uses a fitness-weighted approach, can yield 10-35% better explanations and enable efficient baseline filtering, directly enhancing the fidelity and computational efficiency of your XAI pipelines.

Key insights

Weighted Average Gradients (WG) improves feature attribution by fitness-weighting baselines, outperforming uniform averaging and reducing variance.

Principles

• Baselines for feature attribution should not be treated equally.
• Fitness-weighted aggregation enhances explanation accuracy and stability.
• Explanation methods must satisfy completeness, sensitivity, and implementation invariance.

Method

WG employs an unsupervised perturbation-based fitness function, D_alpha, to assess baseline quality. Weights are inversely proportional to D_alpha, and an O(log n) binary search algorithm efficiently computes D_alpha for each baseline.

In practice

• Filter suboptimal baselines to streamline computation.
• Achieve 10-35% improvement over Expected Gradients.
• Maintain high accuracy while reducing computational cost.

Topics

• Explainable AI
• Feature Attribution
• Integrated Gradients
• Expected Gradients
• Weighted Average Gradients
• Baseline Selection
• Variance Reduction

Code references

• Tamnt240904/weighted_baseline

Best for: Research Scientist, Computer Vision Engineer, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• DAG-MoE: From Simple Mixture to Structural Aggregation in Mixture-of-Experts — DAG-MoE improves Mixture-of-Experts performance by replacing weighted-summation with learned structural aggregation, expanding expert-combination space and enabling multi-step reasoning.
• MLS Part 14: XGBoost Explained — The Powerful Gradient Boosting Algorithm Behind Winning ML Models — XGBoost enhances gradient boosting with second-order optimization and regularization for superior speed and accuracy.
• The Mechanism of Weak-to-Strong Generalization: Feature Elicitation from Latent Knowledge — W2S generalization enables strong models to learn from weak supervision without catastrophic forgetting of other capabilities.
• Correcting Performance Estimation Bias in Imbalanced Classification with Minority Subconcepts — Predicted-weighted balanced accuracy (pBA) corrects evaluation bias in imbalanced classification by accounting for minority subconcept heterogeneity.
• The Double Dilemma in Multi-Task Radiology Report Generation: A Gradient Dynamics Analysis and Solution — Linear scalarization in multi-task RRG fails due to gradient dynamics' "Double Dilemma," solvable by CAME-Grad.
• Adaptive Weighted Averaging — The Adaptive Weighted Averaging strategy identifies the largest value among n options, guaranteeing performance never worse than baselines.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by stat.ML updates on arXiv.org.