Muppet: A Modular and Constructive Decomposition for Perturbation-based Explanation Methods

The topic of explainable AI has recently received attention driven by a growing awareness of the need for transparent and accountable AI. In this paper, we focus on local perturbation-based explanation methods, which currently represent a majority of the post-hoc explainability literature and usage. We have observed a fundamental commonality among these methods, forming the essence of our contribution. We propose a novel methodology to decompose any state-of-the-art perturbation-based explainability approach into four blocks. This decomposition framework offers a concise analysis of existing

Similarities between methods and accelerates the development of new ones and their variants by promoting the reuse of their common blocks. In addition, we provide Muppet: an open-source Python library that offers (i) explainable AI methods to debug and interpret ML models, (ii) standardized API based on our decomposition methodology, facilitating contribution and benchmarking, and (iii) built-in modularity design: enables the decomposition of every method into reusable modules, thus speeding up their implementation. Available soon at https://github.com/euranova/muppet/.

Quentin Ferre, Ismail Bachchar, Hakima Arroubat, Aziz Jedidi, Youssef Achenchabe, Antoine Bonnefoy, Muppet: A Modular and Constructive Decomposition for Perturbation-based Explanation Methods, In Proc. of the International Joint Conference on Neural Networks, July 2025.

Click here to access the paper.