Adding Hidden Nodes to Gene Networks (Extended Abstract)

Bayesian networks are widely used for modelling gene networks. We investigate the problem of expanding a given Bayesian network by adding a hidden node - a node on which no experimental data are given. Finding a good expansion (a new hidden node and its neighborhood) can point to regions where the model is not rich enough, and help locate new, unknown variables that are important for understanding the network. We study the computational complexity of this expansion, show it is hard, and describe an EM based heuristic algorithm for solving it. The algorithm was applied to synthetic datasets and to yeast gene expression datasets, and produces good, encouraging results.