Causal discovery in time series with unobserved confounders

Abstract

Understanding cause and effect relationships is an essential part of the scientific inquiry. There are, however, many circumstances in which the classical approach of controlled experimentation is not feasible. This is the case, for example, for most aspects of Earth's complex climate system. In the first part of the talk we will give a brief introduction into the modern framework of causal inference and causal discovery, which provides alternative methods for learning and reasoning about cause and effect from observational, i.e., non-experimental data. These methods have in recent years been gaining increasing attention from various research fields, for example from the climate and Earth system sciences as well as from the machine learning and artificial intelligence community. We will then present the novel LPCMCI causal discovery algorithm for learning the cause and effect relationships in multivariate time series. This algorithm is specifically adapted to several challenges that are prevalent in time series considered in the climate and Earth system sciences, for example strong autocorrelations, combinations of time lagged and contemporaneous causal relationships, as well as nonlinearities. It moreover allows for the existence of latent confounders, i.e., unobserved common causes, a complication that is faced in most realistic scenarios.