Shapley Image Explanations With Data-Aware Binary Partition Trees

Muhammad Rashid, Elvio G. Amparore, Enrico Ferrari, Damiano Verda

Keywords: Shapley coefficients, Owen approximation, Binary Partition Tree, eXplainable AI

TL;DR: A novel explainable AI method that enhances the visual interpretation of learned representations of images by aligning Shapley value approximations with morphological image features through a data-aware coalition structure.

Abstract:

Extracting a visual interpretation of a learned representation of a machine learning model applied to image data is a relevant task in eXplainable AI (XAI).
Pixel-level feature attributions serve as a valuable tool in this context, as they identify the regions within an image responsible for the classification outcome.
The hierarchical Owen approximation of the Shapley values has proved to be an effective strategy for this task.
However, existing approaches lack data-awareness, leading to poor alignment between the pixel-level attributions and the actual morphological features of the classified image.

This paper introduces ShapBPT, a novel XAI method that computes the Owen approximation of the Shapley coefficients following a data-aware binary hierarchical coalition structure derived from the Binary Partition Tree computer vision algorithm.
By aligning with the morphological features of the image, the proposed method significantly enhances the identification of relevant image regions.
Experimental results confirm the effectiveness of the proposed method.

Primary Area: interpretability and explainable AI

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 6382