Inferring causal directions from observed variables is one of the fundamental problems in artificial intelligence(AI) field. Traditional conditional independence based methods usually learn causal directions by detecting V-structures and return Markov equivalence classes, instead of true causal structures. Most other direction learning methods can distinguish the equivalence classes, but are effective only in the bivariate(or two-dimensional) cases. This paper proposd a new approach for causal direction inference from general networks, based on a split-and-merge strategy. The method first decomposed an n-dimensional network into n induced subnetworks, each of which corresponded to a node in the network. Each induced subnetwork could be subsumed to one of the three substructures: one-degree, non-triangle and triangle-existence structures. It deve-loped three effective algorithms to infer causalities from the three substructures, and learning these induced subnetworks orderly to achieved the whole causal structure of the multi-dimensional network. Experiments show that the method is more general and effective than traditional methods.

[2019-06-05] Printed Article /doc/2017.12.0802/v2