Learning DAGs and Root Causes from Time-Series Data

Many applications produce time-series data: multi-dimensional data measured in regular time steps. Examples include temperature measurements at different sites in meteorology [Yang et al., 2022], stock prices in finance [Kleinberg, 2013, Jiang and Shimizu, 2023], and brain data in medicine [Smith et al., 2011]. A key problem in analyzing time-series data is causal structure discovery, which aims to understand the generation mechanism of such data between nodes and across time [Assaad et al., 2022b, Runge et al., 2023, Gong et al., 2023, Hasan et al., 2023]. On common structural model associates time-series data with directed acyclic graphs (DAGs) that encode how the data in one time step is obtained from prior ones. Our work specifically focuses on learning these DAGs from time-series data [Sun et al., 2023, Gao et al., 2022, Pamfil et al., 2020]. This approach simplifies the broader problem of causal discovery by abstracting away the need for true causal relationships, which often require techniques like interventions. Despite this simplification, DAG learning from time series still poses a challenge due to the complexity of temporal dependencies and the high dimensionality of data.