The article addresses the challenge of accurately and efficiently estimating Shapley values to interpret deep learning predictive models. The key highlights are:
Existing methods for Shapley value estimation, such as regression-based, sampling-based, and structure model-based approaches, suffer from various limitations in terms of accuracy, efficiency, or applicability.
The authors propose EmSHAP, an energy model-based approach that can effectively approximate the expectation of the Shapley contribution function under arbitrary subsets of features.
EmSHAP uses a GRU network to estimate the proposal conditional distribution, which eliminates the impact of feature ordering on the estimation accuracy. A dynamic masking scheme is also introduced to improve the generalization ability.
Theoretical analysis is provided to show that EmSHAP achieves tighter error bounds compared to state-of-the-art methods like KernelSHAP and VAEAC, leading to higher estimation accuracy.
Case studies on medical and industrial applications demonstrate that EmSHAP exhibits enhanced estimation accuracy without compromising efficiency.
A otro idioma
del contenido fuente
arxiv.org
Ideas clave extraídas de
by Cheng Lu,Jiu... a las arxiv.org 04-02-2024
https://arxiv.org/pdf/2404.01078.pdfConsultas más profundas