Disclaimer: Please use other Implementation in Practice!

If you plan on using SHAP-IQ or other interaction algorithms for your own purposes, please have a look at our publicly available python package shapiq!

📄 SHAP-IQ: Unified Approximation of any-order Shapley Interactions

This repository holds the supplement material for the contribution SHAP-IQ: Unified Approximation of any-order Shapley Interactions.

🚀 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 ParameterizedSparseLinearModel
from approximators import SHAPIQEstimator

game = ParameterizedSparseLinearModel(
    n=30, # n of players
    weighting_scheme="uniform", # how the interactions should be distributed over the subset sizes
    min_interaction_size=1, # min size of interactions in the model
    max_interaction_size=20, # max size of interactions in the model
    n_interactions=100, # number of interactions in the model
    n_non_important_features=3 # number of dummy (zero weight) features, which will also not be part of the interactions
)

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

interaction_order = 2

SHAP-IQ to approximate the Shapley Interaction Index

shapiq_sii = SHAPIQEstimator(
    interaction_type="SII", N=player_set, order=interaction_order,  top_order=True
)

sii_scores = shapiq_sii.compute_interactions_from_budget(
    game=game_fun, budget=budget,
    pairing=False, sampling_kernel="ksh", only_sampling=False, only_expicit=False, stratification=False
)

SHAP-IQ to approximate the Shapley Taylor Index

shapiq_sti = SHAPIQEstimator(
    interaction_type="STI", N=player_set, order=interaction_order,  top_order=True
)

sti_scores = shapiq_sti.compute_interactions_from_budget(
    game=game_fun, budget=budget,
    pairing=False, sampling_kernel="ksh", only_sampling=False, only_expicit=False, stratification=False
) 

SHAP-IQ to approximate the Shapley Faith Index

shapiq_FSI = SHAPIQEstimator(
    interaction_type="FSI", N=player_set, order=interaction_order,  top_order=True
)

FSI_scores = shapiq_FSI.compute_interactions_from_budget(
    game=game_fun, budget=budget,
    pairing=False, sampling_kernel="ksh", only_sampling=False, only_expicit=False, stratification=False
)

✅ Validate Experiments

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

Sum of unanimity models (SOUM)

To run the experiemtns on the synthetic model functions (synthetic game functions) we refer to experiment_run_soum.py. There you can specify the complexity of the model functions.

Language Model (LM)

To run the experiments conducted on the language model we refer to experiment_run_look_up.py. We compute interaction indices for a sentiment analysis model. The underlying language model is a finetuned DistilBert version: dhlee347/distilbert-imdb For more information on the movie reviews and value function we refer to precompute_lm.py.

Image Classification Model (ICM)

To run the experiments conducted on the image classification model we refer to experiment_run_look_up.py. The underlying image classifier is a ResNet trained on ImageNet as provided by torchvision.models.resnet18. For more information on the superpixels and value function we refer to precompute_icm.py.

SHAP-IQ = UnbiasedKernelSHAP:

To validate the experiment, that UnbiasedKernelSHAP and SHAP-IQ indeed are the same method, we refer to test_unbiased_vs_shapiq.py.

n-SII efficiency throught sampling:

To validate the claim that SHAP-IQ preserves the efficiency of n-SII throughout the sampling process, we refer to test_n_sii_efficiency.py.

FSI is not s-efficient:

To validate the claim that FSI is not s-efficient, we refer to test_s_efficiency.py.