Framework¶

XAIB is build and aimed to be developed further using complete and defined framework of interpretability properties.

In the recent work Nauta et al. proposes complete evaluation scheme for XAI methods considering interpretability as a multi-faceted concept. They argue that interpretability is not a binary property: it has different aspects that can be measured independently. They propose “Co-12 properties” of interpretability for XAI methods, which are namely: Correctness, Completeness, Consistency, Continuity, Contrastivity, Covariate complexity, Compactness, Compositionality, Confidence, Context, Coherence, and Controllability. For each of them they suggest general ways of quantitatively measure them without concrete implementation suggestions and the overview of how different authors measure them.

Authors define an ontology of interpretability which can serve as a framework for the development of new quality metrics.

XAIB tries not to reinvent the wheel in the field of interpretability and uses this framework as the most complete and recent to this day.