Dempster-Shafer Credal Probabilistic Circuits

Probabilistic circuits are deep, tractable generative models capable of computing various types of exact inferences. However, their traditional specifications do not fully account for epistemic uncertainty. To address this, credal probabilistic circuits were introduced, incorporating a way to manage such uncertainty. We propose a novel framework for learning the structure and parameters of credal probabilistic circuits, leveraging the Dempster-Shafer theory of evidence. Unlike previous credal approaches, the framework handles both discrete and continuous data and allows for the use of multiple classification criteria. We conclude by presenting some preliminary experimental results, demonstrating the performance of the proposed models compared to commonly used probabilistic circuits across a range of classification tasks.