Discovering Time-invariant Causal Structure from Temporal Data

Abstract

Discovering causal structure from temporal data is an important problem in many fields in science. Existing methods usually suffer from several limitations such as assuming linear dependencies among features, limiting to discrete time series, and/or assuming stationarity, i.e., causal dependencies are repeated with the same time lag and strength at all time points. In this paper, we propose an algorithm called the μ-PC that addresses these limitations. It is based on the theory of μ-separation and extends the well-known PC algorithm to the time domain. To be applicable to both discrete and continuous time series, we develop a conditional independence testing technique for time series by leveraging the Recurrent Marked Temporal Point Process (RMTPP) model. Experiments using both synthetic and real-world datasets demonstrate the effectiveness of the proposed algorithm.