SPEX

Sparse Explainer via Fourier transforms using SPEX Kang et al.[1].

from __future__ import annotations

import numpy as np

import shapiq

N_PLAYERS = 8
BUDGET = 200
feature_names = [f"x{i}" for i in range(N_PLAYERS)]

weights = np.array([0.4, 0.3, 0.2, 0.1, 0.05, -0.1, -0.2, -0.3])


def game_fun(coalitions: np.ndarray) -> np.ndarray:
    coalitions = np.atleast_2d(coalitions)
    return (coalitions @ weights) + 0.5 * coalitions[:, 0] * coalitions[:, 1]

Approximate interaction values

approximator = shapiq.SPEX(n=N_PLAYERS, max_order=2, random_state=42)
iv = approximator.approximate(500, game_fun)  # SPEX requires a larger budget
print(iv)

InteractionValues(
    index=k-SII, max_order=2, min_order=0, estimated=True, estimation_budget=384,
    n_players=8, baseline_value=-6.938893903907228e-18,
    Top 10 interactions:
        (np.int64(0), np.int64(1)): 0.5
        (np.int64(2),): 0.20000000000000004
        (np.int64(0),): 0.15000000000000002
        (np.int64(3),): 0.1
        (np.int64(1),): 0.050000000000000044
        (np.int64(4),): 0.049999999999999975
        (): -6.938893903907228e-18
        (np.int64(5),): -0.10000000000000005
        (np.int64(6),): -0.2
        (np.int64(7),): -0.30000000000000004
)

Force plot

iv.plot_force(feature_names=feature_names)

Network plot

iv.plot_network(feature_names=feature_names)

References

[1]

Jae Shin Kang, Liam Butler, Aryan Agarwal, Yusef Erginbas, Ramtin Pedarsani, Kannan Ramchandran, and Bin Yu. SPEX: scaling feature interaction explanations for LLMs. CoRR, 2025. arXiv:2502.13870.