(From my LinkedIn article)
I’m happy to announce that I just released the beta version of Generative Perturbation Analysis (GPA) on GitHub
along with the paper, poster, and presentation slides to be presented at KDD 2023 next week.
GPA is an algorithm for anomaly attribution, whose task is to provide an explanation for an observed deviation in the black-box regression setting. It falls into the category of XAI (explainable AI). For comparison purposes, the repo provides a carefully designed Python implementation of major black- and white-box attribution algorithms:
- LIME [Ribeiro et al. KDD 16]
- Integrated gradient (IG) [Sundararajan et al. ICML 20]
- Expected integrated gradient (EIG) [Deng et al. AAAI 21]
- Shapley values (SV) [Strumbelj & Kononenko KAIS 14]
- Likelihood compensation (LC) [Ide et al. AAAI 21]
Hope it is useful to researchers and practitioners.