Bin Han, Rodney O'Donnell, et al. "Incorporating Prior Information in Causal Discovery"

Most BNs are either constructed by eliciting information from experts ("knowledge engineering") or learned via causal discovery programs (K2, TETRAD, CaMML) from sample data. Neither approach is fully satisfactory in practice. Knowledge engineering is expensive and time-consuming (and requires experts); discovery is often ineffective given small, or noisy, datasets. A hybrid approach is to incorporate prior information elicited from experts in the causal discovery process. We have incorporated five ways of specifying prior information into CaMML, substantially enhancing CaMML's performance. In this study, we compare CaMML with and without prior information to a variety of other programs, including K2 and TETRAD (PC and GES versions). Preliminary results show that CaMML achieves comparable results without prior information, and superior performance with prior information.