POMELO: Black-Box Feature Attribution with Full-Input, In-Distribution Perturbations
Abstract
Model-agnostic explanation methods provide importance scores per feature by analyzing a model’s responses to perturbed versions of the sample to be explained. The explanation’s quality therefore hinges on the made perturbations and, most importantly, suffers if these lead to out-of-distribution samples. Unfortunately, this is the case for the popular LIME explanation method. In this paper, we thus introduce POMELO, an extension to LIME leveraging generative AI for full-input, in-distribution sampling. We define key properties of such samplers: distribution alignment, diversity, and locality. Based on these, we discuss different neural samplers based on normalizing flows and diffusion models. Our results demonstrate that neural samplers outperform traditional perturbation strategies and yield explanations that are better aligned with human intuition.
For further details please consult the conference publication.
The figure above shows the methodology of POMELO. First, the input is segmented. In parallel, a neural sampler generates full-input perturbations of the input. The segment-wise differences and the changes in the softlabels are then used to train a linear surrogate model. The weights of this model serve as the importance scores of the associated segments, mapping to sets of pixels. The critical difference to LIME is that we allow full-input perturbations instead of perturbing random segments. Therefore our method captures big and distant correlations more constistently.
Team
- Luan Ademi
- Maximilian Noppel (Contact)
- Christian Wressnegger
Publication
A detailed description of our work was presented at the 3rd World Conference on eXplainable Artificial Intelligence (XAI) in July 2025. If you would like to cite our work, please use the reference as provided below:
@InProceedings{Ademi2025POMELO,
author = {Luan Ademi, Maximilian Noppel and Christian Wressnegger},
title = {POMELO: Black-Box Feature Attribution with Full-Input, In-Distribution Perturbations},
booktitle = {Proc. of 3rd World Conference on eXplainable Artificial Intelligence (XAI)},
year = 2025,
month = july
}