Necessary and Sufficient Explanations of Multi-Criteria Decision Aiding Models, with and Without Interacting Criteria

Necessary and Sufficient Explanations of Multi-Criteria Decision Aiding Models, with and Without Interacting Criteria

---

연구 분야: Verification

---

---

학회: World Conference on Explainable Artificial Intelligence

---

초록

We are interested in explaining models from Multi-Criteria Decision Aiding. These can be used to perform pairwise comparisons between two options, or classify an instance in an ordered list of categories, on the basis of multiple and conflicting criteria. Several models can be used to achieve such goals ranging from the simplest one assuming independence among criteria - namely the weighted sum model - to complex models able to represent complex interaction among criteria, such as the Hierarchical Choquet Integral (HCI). We consider two complementary explanations of these two models under these two goals: sufficient explanation (a.k.a. Prime Implicants) and necessary explanations (a.k.a. Counterfactual explanations). The idea of prime implicants is to identify the parts of the instance that need to be kept unchanged so that the decision remains the same, while the other parts are replaced by any value. We generalize the notion of information that needs to be kept not only on the values of the criteria (values of the instance of the criteria) but also on the weights of criteria (parameters of the model). For the HCI model, we propose a Mixed-Integer Linear Program (MILP) formulation to compute the prime implicants. We also propose a weak version of prime implicants to account for the case where the requirements of changing the other criteria in any possible way is too strong. Finally, we also propose a MILP formulation for computing counterfactual explanations of the HCI model.

---