Improving Explanation Reliability through Group Attribution

Sangmin Lee, Jongseong Jang, Woohyung Lim

Keywords: Explainable AI, Attribution methods, Group attribution, Attribution reliability

TL;DR: We have proposed the group-wise attribution methods to yield more reliable explanation in understanding a model's prediction

Abstract:

Although input attribution methods are mainstream in understanding predictions of DNNs for straightforward interpretations, the non-linearity of DNNs often makes the attributed scores unreliable in explaining a given prediction, deteriorating the faithfulness of the explanation.
However, the challenge could be mitigated by attributing scores to groups of explanatory components instead of the individuals, termed group attribution. While a group attribution would explain the group-wise contribution more reliably, it does not explain the component-wise contributions so that estimating component-wise scores yields less reliable explanation, indicating the trade-off of group attributions.
In this work, we introduce the generalized definition of reliability loss and group attribution, and formulate the optimization problem of the trade-off with these terms. We apply our formalization to Shapley value attribution to propose the optimization method G-SHAP. We show the effectiveness and explanatory benefits of our method through empirical results on image classification tasks.

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.

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: Yes

Please Choose The Closest Area That Your Submission Falls Into: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)