|
Detecting marginal and conditional independencies between
events and learning their causal structure Host Publication: 12th European Conference, ECSQARU Authors: J. Lemeire, S. Meganck, A. Zimmer and T. Dhollander Publisher: Springer Publication Date: Aug. 2013 Number of Pages: 12 ISBN: 978-3-642-39090-6
Abstract: Consider data given as a sequence of events, where each event has a
timestamp and is of a specific type. We introduce a test for detecting marginal
independence between events of two given types and for conditional independence
when conditioned on one type. The independence test is based on comparing
the delays between two successive events of the given types with the delays
that would occur in the independent situation. We define a Causal Event Model
(CEM) for modeling the event-generating mechanisms. The model is based on
the assumption that events are either spontaneous or caused by others and that
the causal mechanisms depend on the event type. The causal structure is defined
by a directed graph which may contain cycles. Based on the independence test, an
algorithm is designed to uncover the causal structure. The results show many similarities
with Bayesian network theory, except that the order of events has to be
taken into account. Experiments on simulated data show the accuracy of the test
and the correctness of the learning algorithm when assumed that the spontaneous
events are generated by a Poisson process. External Link.
|
