From https://en.wikipedia.org/wiki/Shapley_value, it is possible to understand that direct computation of Shapley values is difficult with their general formula :
$$ \varphi_i(v) = \frac{1}{\text{number of players}} \sum_{\text{coalitions excluding }i} \frac{\text{marginal contribution of }i\text{ to coalition}}{\text{number of coalitions excluding } i \text{ of this size}}$$
Basically beacuse the number of coalitions excluding i grows in complexity with $\sum_{k=1}^{n-1} k!$, where n is the number of variables. Some progresses has been made in the direction of evaluating this sum with Monte-Carlo techniques (as mentionned in https://christophm.github.io/interpretable-ml-book/), but those calculations are still intensive.
In their article (http://papers.nips.cc/paper/7062-a-unified-approach-to-interpreting-model-predictions) Lundberg and Lee proposes two new approaches, relying on SHAP - (these are the Shapley values of a conditional expectation function of the original model) :
- A model agnostic approach basically rewriting the problem as a linear regression problem, which is intuitively less expensive to compute. Basically saying that SHAP is a function of the weight of the model and trying to approximate it.
- A model specific approach. Assuming input independence (which is rarely true ...) they show how to compute SHAP value directly from model weights. Starting with linear models they devise similar relations for NN with usual propagation techniques.
As to which method is used exactly for MLP, I don't know exactly, but the second one seems more appropriate (model specific, exact method).
