5 Discussion
Now that we have covered the available methods in depth, we are in a position to revisit the second stage in our proposed framework, namely, how to select an appropriate Shapley explanation method. For a summary of our recommendations represented as a decision tree, see Figure 5.1.
As noted previously, the first step in the process is for the explainer to specify one or more target explanatory questions aligned with the explainees objectives. For each question, the explainer must then consider the following three questions:
- Is the explainee interested in a model or world-level explanation?
- Is the target explanatory question associative, interventional, or counterfactual?
- What degree of causal information is the explainee able to provide?
To answer the first question, we recommend engaging the explainee to understand where a hypothetical intervention, which captures the essence of their question, occurs. If a model-level explanation is required, then SSV, KernelSHAP, causal-QII, SF, and RSV are all appropriate choices. Due to the availability of open-source implementations, the relative simplicity of the method, and the fact that no auxiliary information must be provided by the explainee, we recommend KernelSHAP.
The second question is only relevant when a world-level explanation is required. For associative questions, we recommend either SRV or ConditionalKernelSHAP. When the number of features and the dataset are relatively small, we recommend SRV because it is an exact method, which limits the number of additional considerations that must be taken into account. For larger datasets or for models with a large number of features, we suggest Conditional KernelSHAP. Again, this recommendation is partially pragmatic due to the availability of an open-source implementation, which allows for different methods of estimating the required conditional distributions.
For interventional and counterfactual model-level explanations, the explainer must be able to provide auxiliary causal information in order for the question to be answerable. One way to identify the degree of causal information available is to engage the explainee in a process of generating a graphical causal model for their application. In the process, it may become clear that the explainee is either unable or unwilling to commit to providing such information. If the explainee can provide only information to develop the limited version of a GCM as required by CSV, then CSV is the clear choice. If a full GCM that is consistent with the available data can be elicited, then CSV, SF, and RSV are all appropriate. However, we recommend using CSV as it requires the least amount of auxiliary information and is more computationally efficient than the other two methods.
In exceptional cases, the explainee may be able to provide a full structural causal model. If this is the case, it may be worth working with the explainee to understand if machine learning is necessary to solve the original problem. If an explainee has sufficient domain expertise to generate an SCM, then it is unclear why the elicited SCM is not being used in lieu of a learned model. Provided the SCM is a valid description of the data generating process, it should be able to generate predictions at least as good as the model. Moreover, the SCM is sufficient for answering associative, interventional, and counterfactual questions that address a different level of explanation entirely. Instead of providing world-level model explanations, the SCM is sufficient to provide world-level explanations world explanations, and specifically, the aspect of the world previously predicted by the model.
It is critical to note that a Shapley explanation may not always be appropriate. This could occur for two reasons. First, the answers to the three aforementioned questions may not have a corresponding Shapley method. This could occur in a situation where the explainee is interested in a world-level counterfactual question, but cannot provide the necessary auxiliary causal information. At this point, the explainer may explore non-Shapley-based explanation methods, or go back to the explainee to identify a different explanatory question that is answerable using a Shapley-based method and aligned with their objectives. Alternatively, the answers to these questions may suggest that an entirely different modeling approach is more appropriate. Whether or not this is feasible largely depends on how the explanations will be used. For example, if explanations are required to satisfy a regulatory requirement for a deployed model, then a better alternative modeling approach is irrelevant.