Workgroup Financial Mathematics
print


Breadcrumb Navigation


Content

Causal structure learning for partially observed multivariate event processes

Oberseminar Finanz- und Versicherungsmathematik


Causal structure learning for partially observed multivariate event processes

Abstract

Structural causal models of event processes imply certain local independencies among the coordinates of the processes. The local independencies form an independence model that can be encoded as a graphical separation model in a directed graph via δ- or μ-separation. If only some of the process coordinates are observed, it is important to understand what can be learned about the causal structure in terms of the local independence model. We recently showed that independence models given by μ-separation in directed mixed graphs are closed under marginalization, and we characterized the Markov equivalence class of a graph. This naturally leads to a causal structure learning algorithm when a local independence oracle is available. We propose a way to replace the oracle by statistical tests of local independence to obtain an empirical learning algorithm. The tests are based on expanding a general intensity
process as a Volterra series of iterated integrals.