Greedy PIG: Adaptive Integrated Gradients

Kyriakos Axiotis, Sami Abu-El-Haija, Lin Chen, Matthew Fahrbach, Gang Fu

Primary Area: optimization

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.

Keywords: feature saliency, feature attribution, feature selection, graph neural networks

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We introduce adaptivity to integrated gradients and show significant improvements on feature attribution/selection and graph compression.

Abstract: Deep learning has become the standard approach for most machine learning tasks. Although its great success is undeniable, interpreting the predictions of deep learning models from a human perspective remains a challenge. In contrast to model training, model interpretability is harder to quantify or pose as an explicit optimization problem. Inspired by the AUC softmax information curve (AUC SIC) metric for evaluating feature attribution methods, we propose a unified discrete optimization framework for feature attribution and feature selection based on subset selection. This leads to a natural adaptive generalization of the path integrated gradients (PIG) method for feature attribution, which we call Greedy PIG. We show that Greedy PIG achieves an extremely high AUC SIC for feature attribution tasks on images, which could also hint at the limitations of this metric for multi-class classification, and we propose a more robust metric. We demonstrate the success of Greedy PIG on a variety of tasks, including image feature attribution, graph compression/explanation, and post-hoc feature selection on tabular data. Our results show that introducing adaptivity is a versatile method for making attribution methods more powerful.

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: 7910