Sum-of-Parts (SOP) Models: Faithful Attributions for Groups of Features

[Paper] [Blog]

Official implementation for "Sum-of-Parts: Faithful Attributions for Groups of Features".

Authors: Weiqiu You, Helen Qu, Marco Gatti, Bhuvnesh Jain, Eric Wong

TODO

• Release updated code - Oct 2nd 2024
• Update arxiv - Oct 5th 2024

Prerequisite

Conda

To set up the environment:

conda create -n sop python=3.10
conda activate sop
pip install -r requirements.txt

Docker

Alternatively, you can use the docker image fallcat/xai:latest

Exlib

No matter which of the above options you chose, you need to install exlib

Usage

Demos

Here we show how to use pretrained SOP and train your own SOP models for ImageNet and CosmoGrid

1. ImageNet
2. CosmoGrid

Evaluation

Here we show how we evaluate. The actual scripts we run are in src/sop/run.

1. ImageNet Accuracy
2. ImageNet Purity
3. ImageNet Insertion Deletion
4. ImageNet Sparsity
5. ImageNet Fidelity
6. Cosmogrid Accuracy and Purity

Citation

@misc{you2024sumofpartsfaithfulattributionsgroups,
      title={Sum-of-Parts: Faithful Attributions for Groups of Features}, 
      author={Weiqiu You and Helen Qu and Marco Gatti and Bhuvnesh Jain and Eric Wong},
      year={2024},
      eprint={2310.16316},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2310.16316}, 
}