kolpaczki/svarm-iq

This repository contains the proposed algorithm SVARM-IQ and the code to rperoduce the experiments of the paper "SVARM-IQ; Efficient Approximation of Any-order Shapley-Interactions through Stratification"

📄 Supplement Material for SVARM-IQ Submission.

This repository holds the code for the paper SVARM-IQ: Efficient Approximation of Any-order Shapley Interactions through Stratification accepted at AISTATS 2024.

🚀 Quickstart

For a quick introduction, we refer to the main.ipynb notebook: Install the dependencies via pip install -r requirements.txt and run the notebook.

Setup

from games import LookUpGame
from approximators import InterSVARM as SVARMIQ

game = LookUpGame(
    data_folder="vision_transformer",   # what explanation task / model to use
    n=16                                # number of image patches for the vision transformer
)

game_name = game.game_name
game_fun = game.set_call
n_players = game.n
player_set = set(range(n_players))

interaction_order = 2

SVARM-IQ to approximate the Shapley Interaction Index

svarmiq_sii = SVARMIQ(
    interaction_type="SII", 
    N=player_set,
    order=interaction_order, 
    top_order=False
)

sii_scores = svarmiq_sii.approximate_with_budget(game=game_fun, budget=budget)

SVARM-IQ to approximate the Shapley Taylor Index

svarmiq_sti = SVARMIQ(
    interaction_type="STI", 
    N=player_set, 
    order=interaction_order,  
    top_order=True
)

sti_scores = svarmiq_sti.approximate_with_budget(game=game_fun, budget=budget)

SVARM-IQ to approximate the Shapley Faith Index

svarmiq_fsi = SVARMIQ(
    interaction_type="FSI",
    N=player_set,
    order=interaction_order, 
    top_order=True
)

fsi_scores = svarmiq_fsi.approximate_with_budget(game=game_fun, budget=budget)

✅ Validate Experiments

To run and validate the same experiments as in the paper the lookup data may need to be first pre-computed. For this we refer to precompute_lm.py, precompute_cnn.py, and precompute_vit.py. We provide some example images and sentences in the games/data folder.

The experiments can be run via the files with the prefix exp_ in the root folder.