-
Notifications
You must be signed in to change notification settings - Fork 0
/
conclusion.qmd
33 lines (31 loc) · 2.13 KB
/
conclusion.qmd
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
# Conclusion
In this work, we introduced a human-centric model explanation evaluation
framework grounded in causal reasoning. Central to this framework is a process
that starts with eliciting well-defined target explanatory questions. We then
revisited the Shapley explanation literature in light of this framework,
highlighting the importance of distinguishing between model and world-level
explanations. Distinguishing between the two and recognizing the type of
explanatory question that is of interest to the explainee allows for a more
principled approach to selecting an appropriate Shapley-based method, or in
some cases, to determining that a Shapley explanation is not appropriate.
In surveying the Shapley explanation literature, a number of potential avenues
for future research emerged. First, there is an open question as to whether a
causal perspective justifies non-Shapley explanation methods in particular
circumstances. In particular, if a fully-specified SCM can be elicited, then
should a machine learning approach for generating predictions have been taken
in the first place? As a second example, if model-level explanations are
desired, what are the benefits of using Shapley values instead of an individual
conditional expectation plot (for local explanations) or a partial dependence
plot (for global explanations)? @zhao_causal_2021 have shown that PDP
and ICE plots have causal interpretations in situations where the complement
set satisfies the backdoor criterion.
The debate over reference distributions within the Shapley explanation
literature has parallels in the counterfactual explanation literature,
highlighting the need for additional unification efforts within XAI.
Specifically, @wachter_counterfactual_2018 introduced the notion of
counterfactual explanations to XAI and advocated for an unconditional
distribution. Subsequent work has been divided on whether a marginal or
interventional approach should be taken. These parallels suggest
that a causal perspective may provide a useful foundation for such unification
efforts. In fact, there is some work already moving in this direction
(@viswanathan_model_2021, @beckers_causal_2022).