Efficient identification of independence networks using mutual information

Conditional independence graphs are now widely applied in science and industry to display interactions between large numbers of variables. However, the computational load of structure identification grows with the number of nodes in the network and the sample size. A tailored version of the PC algorithm is proposed which is based on mutual information tests with a specified testing order, combined with false negative reduction and false positive control. It is found to be competitive with current structure identification methodologies for both estimation accuracy and computational speed and outperforms these in large scale scenarios. The methodology is also shown to approximate dense networks. The comparisons are made on standard benchmarking data sets and an anonymized large scale real life example.