Learning to Infer Causal Graphs from Time-Series Data
A recent paper by Löwe et al describes a novel method that constructs causal graphs from time-series data, which infers causal relationships from observational data. Causal relationships are an important way of studying phenomena, because it gives us information on how to intervene in the world to achieve better outcomes. A researcher studying a simple causality relationship between two variables would not benefit significantly from a graphical representation. However, if the amount of variables increases and the causal relationship between the variables is complex, a graphical representation can greatly aid in the study of causality.
Causal graphs are structurally no different from directed graphs that we have seen in class, the causality results from our interpretation of the directed edges. For example, the diagram below can represent a causal graph if we say that node A is an event that causes events B and C. Conversely, the arrows can also represent one way streets between locations A, B and C, in which case there is no causal relationship.
In previous work and in this paper, the authors sample time-series data to construct many versions of causal graphs. The trick that makes this paper interesting (and very citable) is that it leverages similarities between sampled graphs to optimize over all sampled graphs, whereas previous methods conducted separate optimizations over samples. This implementation significantly improved the performance over previous methods, especially for large sample sizes.
One limitation of this paper is that it can only be applied to time-series data. Using time-series data in causal discovery has the advantage that we can impose some known constraints: events later in time cannot cause events earlier in time. I hope that the authors continue to optimize the construction of causal inference graphs in other data where the time-related causal constraint does not hold.