#ML SHAP (SHapley Additive exPlanations) is a system of meth | Am Neumarkt 😱

#ML

SHAP (SHapley Additive exPlanations) is a system of methods to interpret machine learning models.
The author of SHAP built an easy-to-use package to help us understand how the features are contributing to the machine learning model predictions. The package comes with a comprehensive tutorial for different machine learning frameworks.

- Python Package: [slundberg/shap](https://shap.readthedocs.io/)
- A tutorial on how to use it: https://www.aidancooper.co.uk/a-non-technical-guide-to-interpreting-shap-analyses/

---

The package is so popular and you might be using it already. So what is SHAP exactly? It is a series of methods based on Shapley values.

> SHAP (SHapley Additive exPlanations) is a game-theoretic approach to explain the output of any machine learning model.
>
> -- [slundberg/shap](https://github.com/slundberg/shap)

Regarding Shapley value: There are two key ideas in calculating a Shapley value.
- A method to measure the contribution to the final prediction of some certain combination of features.
- A method to combine these "contributions" into a score.

SHAP provides some methods to estimate Shapley values and also for different models.

The following two pages explain Shapley value and SHAP thoroughly.

- https://christophm.github.io/interpretable-ml-book/shap.html
- https://christophm.github.io/interpretable-ml-book/shapley.html

References:
- Lundberg SM, Lee SI. A unified approach to interpreting model predictions. of the 31st international conference on neural …. 2017. Available: http://papers.nips.cc/paper/2017/file/8a20a8621978632d76c43dfd28b67767-Paper.pdf
- Lundberg SM, Nair B, Vavilala MS, Horibe M, Eisses MJ, Adams T, et al. Explainable machine-learning predictions for the prevention of hypoxaemia during surgery. Nature Biomedical Engineering. 2018;2: 749–760. doi:10.1038/s41551-018-0304-0

---
I posted [a similar article years ago in our Chinese data weekly newsletter](https://github.com/data-com/weekly/discussions/27) but for a different story.