Interpreting Time-Varying Dynamic Bayesian Networks for Earth Climate Modelling

Bayesian networks tend to be considered as transparent and interpretable, but for big and dense networks they become harder to understand. This is the case of non-stationary, and more generally time-varying dynamic Bayesian networks, as the relations change over time and cannot be represented with a single template model. We introduce methods to explain how the model evolves qualitatively over time, and quantify these changes. In addition, we offer a functional implementation for time-varying dynamic Bayesian networks that includes our explainability proposals and some extensions that are targeted to simplify the networks in the specific field of climate sciences.