Shap: Implementation of DeepLift in DeepShap
Is DeepShap(DeepLift + Shapley values) implemented in your current implementation? I am unable to find the computation of DeepLift's multipliers(as given on page 7 of paper) in your current implementation.
Also is this current implementation of DeepShap suitable for Recurrent Neural Networks?
farh33n
>All comments
DeepExplainer uses these multipliers implicitly by overriding the built-in gradient back propagation routine of deep learning libraries. As for the suitability for recurrent networks, it certainly can work and we have seen some results that it seems to work https://arxiv.org/abs/1911.11888, but we would like to extend these tests to our benchmark code in SHAP. Getting this benchmark scaled up is something I am trying to get going in the new year.
slundberg
on 17 Dec 2019
👍3
Was this page helpful?
0 / 5 - 0 ratings
Related issues
shoaibkhanz
路
4Comments
resdntalien
路
3Comments
gabrielcs
路
3Comments
SaadAhmed96
路
3Comments
TdoubleG
路
4Comments
Most helpful comment
DeepExplainer uses these multipliers implicitly by overriding the built-in gradient back propagation routine of deep learning libraries. As for the suitability for recurrent networks, it certainly can work and we have seen some results that it seems to work https://arxiv.org/abs/1911.11888, but we would like to extend these tests to our benchmark code in SHAP. Getting this benchmark scaled up is something I am trying to get going in the new year.