标题： 从时序到同时性迭代因果发现，在存在潜在混杂因素的情况下 显示英文标题

标题： From Temporal to Contemporaneous Iterative Causal Discovery in the Presence of Latent Confounders

摘要： 我们提出了一种基于约束的算法，用于从观测时间序列数据中学习因果结构，在存在潜在混杂因素的情况下。 我们假设一个离散时间、平稳的结构向量自回归过程，包含时间和同时的因果关系。 有人可能会问，时间关系和同时关系是否应该被区别对待。 所提出的算法通过先学习长期时间关系再学习短期关系来逐步优化因果图，其中同时关系最后被学习。 这种学习因果关系的顺序导致所需统计检验的数量减少。 我们通过实证验证了这种减少，并证明与最先进的算法相比，它在合成数据上提高了准确性，并在真实数据上产生了更合理的因果图。 显示英文摘要

摘要： We present a constraint-based algorithm for learning causal structures from observational time-series data, in the presence of latent confounders. We assume a discrete-time, stationary structural vector autoregressive process, with both temporal and contemporaneous causal relations. One may ask if temporal and contemporaneous relations should be treated differently. The presented algorithm gradually refines a causal graph by learning long-term temporal relations before short-term ones, where contemporaneous relations are learned last. This ordering of causal relations to be learnt leads to a reduction in the required number of statistical tests. We validate this reduction empirically and demonstrate that it leads to higher accuracy for synthetic data and more plausible causal graphs for real-world data compared to state-of-the-art algorithms.