Learning gene regulatory networks: instability of constraint-based causal structure learning methods: Marloes Maathuis
Abstract: We considered learning the gene regulatory network of yeast from a high-dimensional observational gene expression data set. We found that constraint-based causal structure learning methods are highly instable, in the sense that their output strongly depends on the ordering of the variables. We studied the various sources of this order-dependence and resolved each one of them. As an alternative solution, we combined causal search methods with sub-sampling methods, where the variable ordering is permuted in each sub-sample. A combination of these two approaches worked best for the yeast gene expression data.