Abstract: Causal discovery from multivariate time-series is a significant and fundamental problem in numerous disciplines.The existing multivariate time-series causal discovery methods learn the causal relations for each individual while some individuals may share the same causal relations;therefore, they may exploit data insufficiently.To this end, this study proposes a collective causal discovery algorithm for multivariate time-series, which is a two-stage algorithm.The first stage measures the similarity of individuals from the perspective of causal relations and clusters the individuals into different groups based on similarity without assigning the number of groups.The second stage involves learning the collective causal relations for each group using variational inference, which sufficiently utilizes the data of individuals in the same group.The experimental result shows that the proposed method outperforms existing methods on simulated data, and the AUC scores are improved by 5%-20%.On real data, the proposed algorithm can separate groups with different causal relations and determine the difference in causal relations for each group, which illustrates the capability of the proposed algorithm in causal discovery and multivariate time-series clustering.

Key words: collective causal discovery, multivariate time-series, causal relations, clustering, variational inference

摘要： 从多元时间序列观测数据中学习多个变量之间的因果关系是许多专业领域中的重要基本问题。现有的多元时间序列因果关系发现方法通常从每个个体的观测数据中学习个体因果关系，没有考虑部分个体之间可能存在相同的因果关系，导致样本利用不足。提出一种面向多元时间序列的群体因果关系发现算法。该算法分为2个阶段：第一阶段基于因果关系对个体之间的相似性进行度量，并把多个个体划分成多个群体，且无须指定群体的个数；第二阶段基于变分推断方法充分利用每个群体内的所有个体数据，从而学习群体因果关系。实验结果表明，该算法在多组不同参数生成的仿真数据上均具有较好的表现，与对比算法相比，AUC评分提升了5%~20%。在真实数据集中，该算法能够较好地区分具有不同因果关系的群体，并且能够学习到不同群体之间不同的因果关系，表明算法不仅具有因果关系发现能力，而且还具有多元时间序列聚类能力。

关键词: 群体因果发现, 多元时间序列, 因果关系, 聚类, 变分推断

CLC Number: 

• TP301.6