Consistent and Truthful Interpretation with Fourier Analysis

Yifan Zhang, Haowei He, Yang Yuan

Keywords: AI Interpretability

TL;DR: We find that the previous attribution methods are not consistent with neighborhood predictions, and introduce a new framework with an efficient algorithm to support consistency.

Abstract:

For many interdisciplinary fields,
ML interpretations need to be consistent with \emph{what-if} scenarios related to the current case, i.e., if one factor changes, how does the model react?
Although the attribution methods are supported by the elegant axiomatic systems, they mainly focus on individual inputs,
and are generally inconsistent.
To support what-if scenarios, we introduce a new objective of consistency based on a notion called truthful interpretation. Towards this objective,
we apply Fourier analysis of Boolean functions to get
consistency guarantees.
Experimental results show that
for neighborhoods with various radii,
our method achieves $2$x - $50$x lower inconsistency compared with the other methods.

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: Theory (eg, control theory, learning theory, algorithmic game theory)

Community Implementations: [ 1 code implementation](https://www.catalyzex.com/paper/arxiv:2210.17426/code)