Causal discovery for time series with latent confounders

TL;DR

This paper evaluates the LPCMCI algorithm for causal discovery in multivariate, autocorrelated time series with latent confounders, highlighting its strengths and limitations in detecting various types of dependencies.

Contribution

The study provides an empirical assessment of LPCMCI's performance on complex time series data with unobserved variables, demonstrating its relative effectiveness and challenges.

Findings

LPCMCI outperforms random algorithms significantly.

Performs best on auto-dependencies.

Struggles most with lagged dependencies.

Abstract

Reconstructing the causal relationships behind the phenomena we observe is a fundamental challenge in all areas of science. Discovering causal relationships through experiments is often infeasible, unethical, or expensive in complex systems. However, increases in computational power allow us to process the ever-growing amount of data that modern science generates, leading to an emerging interest in the causal discovery problem from observational data. This work evaluates the LPCMCI algorithm, which aims to find generators compatible with a multi-dimensional, highly autocorrelated time series while some variables are unobserved. We find that LPCMCI performs much better than a random algorithm mimicking not knowing anything but is still far from optimal detection. Furthermore, LPCMCI performs best on auto-dependencies, then contemporaneous dependencies, and struggles most with lagged dependencies. The source code of this project is available online.