I just recently started to work again on Dynamic Bayesian Networks, and looking at the literature, it seems to me that they were popular in the '00, but less widely in use now, for what it concerns data assimilation and modeling stochastic processes/time series. Is it a false impression, due to the fact I haven't used them for quite some time, or did they actually fall out of grace recently? If so, which approaches are more common in statistical practice now? ARIMAX models? RNNs?
Also, what is the reason why they're less popular? Do they have inherent limitations? For example, maybe the Markov property implies that using a DBN to model problems where high order lags are important, forces us to use too complex a model, with respect to other approaches. Or there may be other reasons I don't see right now.
