author ：18 the cyl

date ：2021-08-15

The paper ：《Shapley Explanation

Networks》

Periodical ：ICLR

One 、 Prerequisite knowledge ：Shapley Value

Shapley Value Originated in game theory ,n People cooperation , Co created v(N) The value of , How to evaluate everyone's value ？ By calculating each
One of the Shapley value.
The formula ：


characteristic ：

1. You can assess the importance of a set of features
2. The calculation is complicated by the exponential level
3. You can do a linear transformation , Find a function shapley value , If the function is transformed linearly , You can use the original
   shapley The value is linearly transformed to find a new shapley value .

Two 、 background

The interpretability of neural network is an important problem to be solved , Understanding the importance of each feature is one direction to improve interpretability .
Proposed in game theory Shapley Value is a way to evaluate the importance of constituent elements , also shapley Values have been applied to machine science
Learning field , But there are two problems ：
（1） because Shapley The characteristics of the value calculation method itself , Its time complexity is exponential ;
（2） The current analysis methods are “ After the event ” Of , That is, the model has been built , Calculate the model shapley value , In this way, the consumption
between , Second, to get shapley Values are only used to analyze , The model itself is not modified .
The goal is
1） Reduce the time complexity by approximation
2） Use Shapley Value guided regularization 、 prune , Modify the model

3、 ... and 、Shapley Explanation Networks