Another paper (but quite complicated): https://arxiv.org/abs/2007.04131 Submitted on 8 Jul 2020 (v1), last revised 17 Aug 2021

Derek

The paper walks through existing model-agnostic interpretation techniques (partial dependence plots (PDP), permutation feature importance (PFI) and Shapley values) and warns against the risky points to take care before trusting the explanations. These points could be captured via XE Evaluation forms typically in iSee and the processing of answers can be mapped to one of the aim for reducing such risks : check that the model is applied with the right context, models can not be generalised or generalising their usage to other sectors/scenarios can lead to a quality damage; check the features dependancy- some being scenario specific features from data set. On particular risk “making unjustified causal interpretations” is difficult to identify

global evaluation and local evaluations do not have the same goal but they tend to share the same pitfalls.