Learning Generalized Hypergeometric Distribution (GHD) DAG models

We introduce a new class of identifiable DAG models, where each node has a conditional distribution given its parents belongs to a family of generalized hypergeometric distributions (GHD). a family of generalized hypergeometric distributions (GHD) includes a lot of discrete distributions such as Binomial, Beta-binomial, Poisson, Poisson type, displaced Poisson, hyper-Poisson, logarithmic, and many more. We prove that if the data drawn from the new class of DAG models, one can fully identify the graph. We further provide a reliable and tractable algorithm that recovers the directed graph from finitely many data. We show through theoretical results and simulations that our algorithm is statistically consistent even in high-dimensional settings ($n >p$) if the degree of the graph is bounded, and performs well compared to state-of-the-art DAG-learning algorithms.