Comparing Instance Attribution to k-Nearest Neighbors

This is the code for the paper A Comparison of Instance Attribution Methods. All third party code is publicly available except for the pre-trained ExPred model, contact Avishek Anand for more information. This is part of the 2023 TuDelft Research Projec (https://github.com/TU-Delft-CSE/Research-Project).

Usage:

1. Pull FastIF, TracIn, and the pre-trained ExPred model, and import into designated files folders.
2. From FastIF only experiments and influence_utils are required. Use the included nn_influence_utils.py. Minor changes to other files may be necessary.
3. In src/TracIn put the files alongside main_tracin.py
4. Install the required packages from the requirements.txt by pip install -r requirements.txt
5. In main.py adjust constants as needed and run.

Not that depending on your hardware you may have to change the k in main.py.

Structure of the repo

├── dataset: the eraser FEVER datasets and custom dataset utils
├── expred: the pretrained expred model (not publicly available)
├── fastif: fastif from https://github.com/salesforce/fast-influence-functions
├── README.md: this file
├── tracin: tracin from https://github.com/frederick0329/TracIn
└── main.py: main execution file