Learning Non-Stationary Dynamic Bayesian Network Structure from Data Stream

Dynamic Bayesian network (DBN) is a useful model for identifying conditional dependencies in time-series streaming data. Non-stationary Dynamic Bayesian network (nsDBN) is a special DBN in which the conditional dependence structure of the underlying data-generation process is permitted to change over time. The current nsDBN structure learning methods can handle abrupt concept drift in streaming data well, but not gradual concept drift. In this paper, we propose a nsDBN learning method that can handle gradual concept drift well. This method first learns the initial DBN structure using local search and global optimization algorithms and then updates the transition networks by continuously searching the structure space with one edge change for each step. The Bayesian-Dirichlet equivalent (BDe) metric is extended for efficient calculation and local searching sub-tasks are executed in parallel from historical local optimized structures. The experimental evaluations show that this method had better precision than other popular methods when learning nsDBN from stream data with gradual concept drift.