中文

从时间序列因果图抽象中总效应的可识别性

统计理论 2025-02-18 v8 人工智能 统计理论

摘要

我们研究在实践常见情形下,仅能获取真实因果图抽象时,从观测时间序列中干预总效应的可识别性问题。此处我们考虑两种抽象:扩展摘要因果图,其合并所有滞后因果关系但区分滞后与瞬时关系;以及摘要因果图,其不给出因果关系之间滞后的任何指示。我们证明总效应在扩展摘要因果图中总是可识别的,并给出摘要因果图中可识别性的充分条件。我们进一步提供了在可识别时用于估计总效应的调整集。

关键词

引用

@article{arxiv.2310.14691,
  title  = {Identifiability of total effects from abstractions of time series causal graphs},
  author = {Charles K. Assaad and Emilie Devijver and Eric Gaussier and Gregor Gössler and Anouar Meynaoui},
  journal= {arXiv preprint arXiv:2310.14691},
  year   = {2025}
}

备注

Correction to original version published at UAI 2024. The published version contains an error in Condition 2(c) in Theorem 2, which is corrected in this version