ShapBPT: Image Feature Attributions using Data-Aware Binary Partition Trees
Date:
🧠 Overview
This talk introduces ShapBPT, a novel method for image feature attribution in Explainable Computer Vision (XCV).
While existing hierarchical Shapley-based methods rely on fixed partitions, they fail to capture the intrinsic multiscale structure of images, leading to inefficient and less meaningful explanations.
🔍 Key Idea
ShapBPT integrates:
- Shapley/Owen values
- Binary Partition Trees (BPT)
to create a data-aware hierarchical representation of images.
👉 This allows explanations to follow natural image structures, rather than arbitrary grids.
⚙️ Key Contributions
- Data-aware hierarchical explanation framework using BPT
- Improved efficiency in Shapley value computation
- Better alignment with object boundaries and semantics
- Validated through experiments and a 20-subject user study
🔗 Resources
Paper arXiv Code Tests PyPI Poster Talk
🧪 Method Overview
ShapBPT explains model predictions by assigning attribution scores to image regions through a hierarchical decomposition:
🔄 Explanation Process
📊 Results & Comparisons
🎯 Impact
ShapBPT demonstrates that:
Explanations should follow data structure, not arbitrary partitions.
It provides a more efficient, stable, and human-aligned approach to explain deep vision models.
🏷️ Keywords
Explainable AI · Shapley Values · Binary Partition Trees · Feature Attribution · Computer Vision · XAI · AAAI 2026