Year: 2025
Type: article
Abstract: Shapley Values (SVs) are concepts established for explaining black-box machine learning models by quantifying feature contributions to model predictions. Since their computation has exponential complexity in the number of features, a variety of approximated approaches to SVs estimation have been proposed. The state-of-the-art neural SVs estimator, i.e., FastSHAP, advances sampling-based methods by first training a supervised surrogate model to learn the conditional expectation of the original model given every feature subset and then generating SVs estimates through a separate network using weighted least squares. While ensuring fast SVs inference, this approach requires a significant training time, thus becoming unsuitable for scenarios where the black-box model has to be frequently retrained. To address this limitation, we propose LIGHTNINGSHAP, a cost-effective neural network-based estimator that jointly computes conditional expectations and SVs estimations to reduce training cost. The unified network learning process minimizes feature-level estimation errors while preserving SVs efficiency. Experiments run on both dynamic and static tabular datasets show that LIGHTNINGSHAP achieves a 25%-60% speedup in overall computational time (i.e., the sum of training and inference times) on medium-large datasets compared to existing neural SVs estimators as well as lower or comparable estimation errors. Furthermore, the results obtained on image datasets indicate that LIGHTNINGSHAP yields a 30%-55% speedup while preserving explanation quality.
Authors: Davide Napolitano, Luca Cagliero
Institution: Politecnico di Torino
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
Funders: European Commission, NextGenerationEU
Award: G4508289328
OpenAlex ID: W4417438479