Causal Discovery from Heterogenous Multivariate Time Series

This paper explores three key aspects of causal discovery from heterogeneous time series. First, it introduces a method that uses (conditional) mutual information to determine (conditional) independence among diverse qualitative and quantitative variables, alongside a novel local permutation test. Next, the paper presents a new algorithm for identifying event-based causal relations in threshold-based IT systems for root cause analysis. This method is effective when root causes are not causally related, and an extension involving agent intervention is proposed to address this limitation. Both the algorithm and its extension utilize causal discovery from offline data and subgraph traversal for new anomalies in online data. Finally, the paper addresses time series with multiple piecewise consistent regimes, each having distinct causal mechanisms. The proposed method segments the time series into appropriate regimes and identifies the correct window causal graph, capturing both instantaneous and lagged connections within each regime. Experiments with synthetic data confirm the effectiveness of the proposed methods.