Contrastive XAI
Explainable AI (XAI) tools often show us the main features that lead to a positive prediction. SHAP explains a prediction with features having the highest marginal contribution, for example. LIME will explain the key features that locally had the highest values in the vicinity of an instance to prediction. In general, we look for the key features that push a prediction over the true or false boundary of a model.
However, IBM Research has come up with another idea: explaining a prediction with a missing feature. The contrastive explanations method (CEM) can explain a positive prediction with a feature that is absent. For example, Amit Dhurandhar of IBM Research suggested that a tripod could be identified as a table with a missing leg.
At first, we might wonder how we can explain a prediction by focusing on what is missing and not highlighting the highest contributions of the features in the instance. It might seem puzzling. But within...
