Abstract:
|
Directed Acyclic Graphs (DAG) capture causal relations between random variables in various fields. However, their reconstruction (estimation) from observational data is a challenging task due to the computational hardness of the learning task. We study this estimation problem when prior information exists on partial orderings of sets of nodes. Specifically, it is assumed that the nodes are partitioned in different sets for which ordering information exists. We develop a framework for a high-dimensional sparse regime, that combines penalized regression to delineate relationships amongst nodes in different sets, with the popular PC-algorithm that identifies the skeleton of the graph within each set. In the final step, we combine the results from the regression and PC-algorithm steps and develop an adjustment step to eliminate redundant edges. The framework is evaluated on simulated data sets from the DREAM3 competition.
|