dPGM

dPGM is an acronym that can denote more than one concept in probabilistic modeling, and its exact meaning depends on the context. Two common interpretations are dynamic probabilistic graphical models and diffusion-based probabilistic graphical models.

Dynamic probabilistic graphical models refer to frameworks for modeling temporal processes with probabilistic structures. They encode dependencies among random variables across time slices, extending conventional graphical models to sequences. This enables reasoning about evolving states and time-dependent relationships. Inference in such models can be performed with methods akin to belief propagation, Kalman filtering, or particle filtering, and may be exact or approximate depending on the model’s structure.

Diffusion-based probabilistic graphical models integrate diffusion processes into a graphical-model framework to capture data generation or noise processes with temporal or structural aspects. In these approaches, a forward diffusion process gradually adds noise and a learnable reverse process aims to denoise, supporting tasks such as data generation or reconstruction within a probabilistic setting. Inference often employs variational techniques, score-based methods, or other optimization-based approaches.

Because the term dPGM is used across different disciplines, its precise definition varies between publications and applications. Some authors use it as a generic shorthand for any probabilistic graphical model with a dynamic or diffusion component, while others define specific architectures or variants. When encountering the term, it is helpful to consult the source to determine the intended meaning.

See also: probabilistic graphical models; dynamic Bayesian networks; diffusion models; state-space models.