初窥Shapley Values

Shapley Values

Definition

从同事那里了解到Python有个机器学习相关的库-SHAP，可以用来做模型特征重要性相关的探索。因此花点时间了解一下SHAP。既然提到SHAP，那么肯定要了解一下其背后的Shapley Value。

从Interpretable machine learning 5.10 SHAP找到一段相关的Shapley Values的定义。

A prediction can be explained by assuming that each feature value of the instance is a “player” in a game where the prediction is the payout. Shapley values – a method from coalitional game theory – tells us how to fairly distribute the “payout” among the features.

将整个模型训练作为一局博弈，每个特征当做一个博弈者，而预测结果则是最终的结果。Shapley Values是一种用来把最终的结果分摊到每个博弈者身上的方法。换句话说，也就是用来判断每个特征对预测结果做出了多大的贡献。

顺带对原文中提到的coalitional game theory了解了一下。Coalitional game theory，或Cooperative game theory，中文直译是合作博弈论，又称正和博弈。是指由于受到可能的外力原因，博弈者们组成联盟的形式，在集团之间的博弈对抗。这与传统的博弈论不同，传统博弈论中博弈者通常被假设为所有决策都是自我主导的，对盟友不存在原谅的可能性的（用人话说，博弈者不可能形成联盟或任何合作机制都必须为自我履约契约）。

回到Shapley Values这个话题，乍一看，Shapley Values的概念不就和Linear Model一个意思嘛？通过每个特征的系数来直观判断特征重要性。但是Linear Model的概念仅适用于简单模型，如果涉及到LGBM之类的复杂模型，就无法适用了。

Example for General Idea

原文中提到一个很有意思的例子。训练一个模型用来预测房价，相关的变量有park，cat，area，floor。如果需要计算cat（banned or allowed）对house price的影响，那么则只需要抽取其余变量相同，只有cat变量不同的样本，计算样本平均贡献-average marginal contribution，即可知道cat变量对house price的影响。 但是这样的计算会存在一个问题，为了达到控制变量的目的，应当对每个变量的所有可能值做排列组合.

Axioms of Shapley Value

Symmetry

i and j are interchangeable relative to v if they always contribute the same amount to every coalition of the other agents.

Which is,
for all S that contains neither i nor j, v ( S ∪ { i } ) = v ( S ∪ { j } ) . v(S\cup\{i\}) = v(S\cup\{j\}). v(S∪{ i})=v(S∪{ j}).

In addition, they should receive the exact same shares/payments.

That is,
for any v, if i and j are interchangeable, then ${\psi }_i(N,v)={\psi }_j(N,v)$

Dummy Players(free rider)

i is a dummy player if the amount that i contributes to any coalition is 0.
Which is,
for all S : v ( S ∪ { i } ) = v ( S ) . S: v(S \cup \{i\}) = v(S). S:v(S∪{ i})=v(S).

Dummy players should receive nothing.
That is,
for any v, if i is a dummy player, then ${\psi }_i(N,v)=0$

Additivity

If we can separate a game into two parts, $v=v_1+v_2$, then we should be able to decompose the payments.

For any two $v_i$, and $v_2,{\psi }_i(N,v_{}$