A Query Language for Explainability

In recent years, the demand for methods to explain individual predictions made by machine learning models has grown significantly. This has led to the emergence of various explanation frameworks and scoring techniques aimed at justifying a model’s classifications. However, rather than struggling with the increasing number of these approaches, an alternative is to draw inspiration from traditional database practices and develop a declarative query language tailored for interpretability tasks. Such a language would enable users to specify and test their own explainability queries effectively. Logic emerges as a natural choice for this purpose, given its well-defined syntax and semantics, along with numerous tools available to analyze its expressiveness and the complexity of query evaluation. In this talk, we will explore recent efforts in designing a logical framework for model interpretability.