Causal discovery task (CDT) in time series become considerable challenging when encountering missing data, as certain crucial information is lost. Therefore, it is necessary to perform missing data imputation task (MDT) to provide more information support for CDT. Essentially, CDT and MDT are two mutually facilitating tasks, because each contains beneficial information for the other. However, most existing studies omit this aspect and treat two tasks independently, suffering from an information crack problem, i.e., potential complementary information within CDT and MDT cannot be sufficiently interacted and exploited, thus inevitably bringing bias and noise to both tasks. Thus, we propose BREAD, a BidiRectional mEta-learning-guided cAusal Discovery framework, to overcome information crack problem from a new standpoint (i.e., bidirectional interaction links between CDT and MDT). Technically, we introduce meta-learning to refine key information characterizing variable influences from causalities and fuse it into MDT, while refining key information characterizing temporal context knowledge from imputation data and fusing it into CDT. Built on it, we can bridge CDT and MDT to realize their bidirectional interactions, thus simultaneously promoting the performance of both tasks. Besides, to improve reliability and robustness of imputation data, we introduce contrastive learning and infuse prior characteristic knowledge to supplement high-quality self-supervised signals. Experiments with various missing mechanisms are performed on simulation and real datasets to verify the superiority of BREAD.

中文翻译：

---

当因果关系遇到缺失数据时：通过双向元学习融合关键信息，在时间序列中桥接因果发现和插补


当遇到缺失数据时，时间序列中的因果发现任务 （CDT） 变得相当具有挑战性，因为某些关键信息会丢失。因此，有必要执行缺失数据插补任务 （MDT） 来为 CDT 提供更多信息支持。从本质上讲，CDT 和 MDT 是两项相互促进的任务，因为每项任务都包含对彼此有益的信息。然而，现有的研究大多忽略了这一方面，独立地处理两项任务，存在信息破解问题，即 CDT 和 MDT 内部潜在的互补信息无法得到充分的交互和利用，从而不可避免地给这两项任务带来偏差和噪声。因此，我们提出了 BREAD，一个 BidiRectional mEta 学习指导的 cAusal Discovery 框架，从新的角度（即 CDT 和 MDT 之间的双向交互链接）克服信息破解问题。从技术上讲，我们引入了元学习来提炼表征因果关系的可变影响的关键信息并将其融合到 MDT 中，同时从插补数据中提炼表征时间上下文知识的关键信息并将其融合到 CDT 中。在此基础上，我们可以桥接 CDT 和 MDT 以实现它们的双向交互，从而同时促进这两个任务的执行。此外，为了提高插补数据的可靠性和鲁棒性，我们引入了对比学习并注入先验特征知识来补充高质量的自监督信号。在仿真和真实数据集上对各种缺失机制进行了实验，以验证 BREAD 的优越性。