A fast dynamic Bayesian network algorithm for structure learning using a time-lagged partial correlation matrix

We propose a novel dynamic Bayesian network algorithm for structure learning using Gaussian Bayesian network analysis and a time-lagged version of the partial correlation matrix. The proposed algorithm overcomes common issues in standard structure learning such as the finding of cycles, time consumption, and computational load. The proposed algorithm is composed of four rules where an initial structure is found using time-lagged partial correlation, followed by a series of steps that evaluate network patterns using the Bayesian information criterion score. We test the accuracy of our method by a simulated multivariate autoregressive network, and also by analysing Saccharomyces Cerevisiae gene expression data.