We develop a penalized likelihood estimation framework to estimate the
structure of Gaussian Bayesian networks from observational data. In contrast to
recent methods which accelerate the learning problem by restricting the search
space, our main contribution is a fast algorithm for score-based structure
learning which does not restrict the search space in any way and works on
high-dimensional datasets with thousands of variables. Our use of concave
regularization, as opposed to the more popular $\ell_0$ (e.g. BIC) penalty, is
new. Moreover, we provide theoretical guarantees which generalize existing
asymptotic results when the underlying distribution is Gaussian. Most notably,
our framework does not require the existence of a so-called faithful DAG
representation, and as a result the theory must handle the inherent
nonidentifiability of the estimation problem in a novel way. Finally, as a
matter of independent interest, we provide a comprehensive comparison of our
approach to several standard structure learning methods using open-source
packages developed for the R language. Based on these experiments, we show that
our algorithm is significantly faster than other competing methods while
obtaining higher sensitivity with comparable false discovery rates for
high-dimensional data. In particular, the total runtime for our method to
generate a solution path of 20 estimates for DAGs with 8000 nodes is around one
hour.

本文提出了一种基于得分的结构学习快速算法的罚似然估计框架，该快速算法不受限制地处理高维数据集，并使用凹规则化以解决估计问题的固有不可识别性。其中，高斯贝叶斯网络是重点研究领域之一。