Thread by @jeremystan

1/ shapley values (and the Python shap package) have become an *integral* part of my machine learning methodology

I’m surprised by how many people still don’t know about them
github.com/slundberg/shap

2/ They allow you to distribute credit for the prediction of an arbitrarily complex machine learning model to the underlying features

They resulting values are additive, and locally explain how the model behaves

3/ At a minimum, they are a better version of common variable importance methods (GBM importance)

They give a distribution of how a variable impacts the prediction, and the direction

4/ This provides crucial feedback for data gathering, feature engineering, and diagnosing model performance issues

5/ But arguably it’s the explanations at an observation level that are most compelling.

They can be integrated into applications and products to help explain why a specific prediction was made

6/ Fuethermore, I think there are many creative applications / products / businesses that they will enable.

For example; they can be used to value data, which could form the basis of marketplaces
arxiv.org/abs/1902.10275

7/ I also think they could be transformative for advertising attribution solutions

8/ I first discovered Shapley values at the NeurIPS conference in 2017.

See the second section in this writeup:
tech.instacart.com/3-nips-papers-…