ProbExplainer: A Library for Unified Explainability of Probabilistic Models and an Application in Interneuron Classification

There are a multiplicity of libraries that implement Bayesian networks and other probabilistic graphical models. However, none of them is dominant, making it hard for users to deploy these models and build over them new functionality, such as new explainability algorithms that are specific for probabilistic models. We provide a common user interface called ProbExplainer for all of them over which algorithms are implemented, leaving to the user the relatively simple task to wrap up the specific implementation. We apply this library in an interneuron classification problem, a domain that is characterised by little expert agreement on labelling. We seek to study feature relevance through map-independence, an explainability method for Bayesian networks. The different expert models agreed that bigger subsets of unobserved features tend to be more relevant, the expert models are divided by whether the columnarity of an interneuron is irrelevant and in general the probability of a new observation changing the classification of its scenario is relatively low.