Learning dynamic bayesian network structures from data

摘要: Dynamic Bayesian networks (DBNs) are graphical models to represent stochastic processes. This dissertation investigates the use of DBNs predict patient outcomes based on temporal data, effectiveness nonstationary multivariate time series and assumptions parametric nature along with two related hypotheses: (1) Given assumption that dataset was generated by stationary first-order Markov processes, patient-specific DBNs, each which a single patient, would mortality more accurately than model an entire population. (2) The predictive performances improve relaxing assumptions. Both hypotheses were tested datasets: A 6704 intensive care unit patients through process simulation. The not supported results evaluated receiver operating characteristics analysis. In light this evidence, new class is called dynamic simple Bayes (DSB) models, developed in dissertation. DSB approach further restricts set conditional independence assumptions; is, all variables any period t conditionally independent given next + 1. Unlike conventional arcs direction flow. Test suggest superior predicting next-day future data. The imply restrictions (e.g., orders probability distributions, or independencies between variables) may lower be preferred baseline for modeling large sample space relatively small size.