Novel DBN Structure Learning Method Based on Maximal Information Coefficient

摘要

Dynamic Bayesian Network (DBN) is a mainstream approach to modeling various biological networks including the gene regulatory network (GRN). For such DBN models that consist only of inter-timeslice arcs, most current methods for learning it employ either a score and search approach or Markov chain Monte Carlo (MCMC) simulation, both of which ignore the structural constraints of DBN models. These structural constraints were first applied to translate the structure learning problem into discovering associations among variables, and then a new method was presented to obtain inter-timeslice arcs. This method was based on maximal information coefficient (MIC). Experiment results showed that the proposed MIC-based method outperformed MI-based, MCMC, and K2 algorithm methods on the quality of learned structure.