Causal Discovery Methods for Climate Networks

Abstract

This paper suggests new methods for the development of network models in climate research. Current climate networks, first introduced in 2004 by Tsonis and Roebber, define network edges based on correlation of node pairs, resulting in a correlation network. The key idea of this paper is to introduce techniques from causal reasoning to derive climate networks, specifically constraint based structure learning. This approach is expected to yield networks that better represent the causal connections in the network, by containing less edges and with all causal pathways still present. The anticipated advantage of a network with less edges is a more manageable model size that makes it easier to gain new insights about causal relationships in the climate system.

The goal of this paper is to provide researchers in the climate area with an intuitive understanding of the causal discovery process, specifically of constraint based structure learning. We review the basic principles of constraint based structure learning, namely how cause-and-effect relationships of variables can be learned from observational data using conditional independence tests. Tutorial-style examples illustrate this process. Finally, we review available algorithms and software packages from other disciplines that can be applied to derive climate networks.

There are no simulation results provided in this paper (work in progress), thus we do not yet know how much reduction is achieved through this method compared to existing methods. However, applications of similar techniques for protein interaction modeling has yielded tremendous savings, making it possible to gain significant understanding of causal pathways from the obtained network graphs.