Causal graph inference
Simona Poilinca, Jhanak Parajuli, Giuseppe Thadeu Freitas de Abreu · 2015
We provide a framework to infer causal relationships in a system of multivariate, stochastic, delayed signals, with application to their prediction. First we address the dimensionality problem in information causality estimation and propose a method to improve the efficiency of calculations by retaining only the most essential components. The directed information between pairs of signals are then used to obtain a maximum spanning tree that captures the strongest causal relationships. Second, causal conditional information is applied to account for further dependencies and obtain the causal graph. Finally, based on this structure, we use delay estimation to accurately predict child signals.