Systems and methods for causal inference in network structures using belief propagation

摘要

Systems and method for perturbing a system include obtaining directed acyclic/cyclic graph candidates {G_I, . . . , G_N} for the system. Each G_i in {G_j, . . . G_N} includes a causal relationship between a parent and child node. {G_I, G_N} demonstrate Markov equivalence. Observed data D is obtained for the nodes. For each respective G_i, the marginal probability of a parent node x_i in G_i is clamped by D while computing a distribution of marginal probabilities for a child node y_i, by Bayesian network or Dynamic Bayesian network belief propagation using an interaction function. The observed distribution for the child node y_i, in D and the computed distribution of marginal probabilities for the child node y_i are scored using a nonparametric function, and such scores inform the selection of a directed/cyclic graph from {G_I, . . . , G_N}. The system is perturbed using a perturbation that relies upon a causal relationship in the selected directed acyclic/cyclic graph.