Comparing with CEM
The Contrastive Explanation Method (CEM) is similar to both anchors and counterfactuals since it explains predictions using what is present (such as anchors) and absent (such as counterfactuals). It calls what is present Pertinent Positives (PPs) and what is absent Pertinent Negatives (PNs). However, the difference is that PPs are qualified as being minimally and sufficiently present to predict the same class. Likewise, PNs are minimally and necessarily absent to predict the opposite class. Therefore, CEM works best with continuous and ordinal features because it expects to subtract from features until it reaches the desired outcome. For this reason, it doesn't know how to deal with non-monotonic continuous, non-ordinal, categorical, or even binary, features, for that matter, and our recidivism dataset only has this kind of feature! Admittedly, this chapter's example doesn't make for an ideal CEM use case. We will touch on CEM in subsequent chapters...
