Abstract:
|
We consider fitting a Bayesian network model to the data where the nodes are mixed with continuous and discrete variables. Most existing approaches either assume that all nodes follow a Gaussian distribution or all nodes follow a multinomial distribution, which limits their applicability to many real-world problems. In this paper, we propose a Gaussian-Probit Bayesian network model with a penalized maximum likelihood estimation using a blockwise coordinate descent (BCD) algorithm. We demonstrate that causal relationships among nodes in a sparse network can be effectively predicted, and the prediction accuracy can be further improved using a consensus network obtained from multiple predictions. The proposed methods were applied to The Cancer Genome Atlas data to study the (epi)genetic pathways that underlie ovarian cancer progression.
|