Interactive Exploration of Comparative Dependency Network Learning

Comparative dependency network learning is a growing field of research, especially in systems biology. Domain scientists would like to discover patterns of variable dependency that are conserved across conditions or discover pathways that are disrupted due to disease. In machine learning, multitask graphical structure learning algorithms have been developed to help solve this problem by learning network models from multiple related datasets. These algorithms typically have regularization hyper-parameters that have the effect of reducing the number of spurious edges learned and the number of spurious differences learned. We propose a mechanism to allow the end-user to control these regularization hyper-parameters in real-time to interactively explore the huge space of potential dependency network solutions. This is a critical element of a visualization system that enables domain scientists to discover interesting patterns in multivariate data. Yet, this is a computationally challenging endeavor as complex models must be learned in real-time and, additionally the number of differences learned in each network and the number of differences between them must be translated by the machine learning algorithm into the correct change in the setting of the hyper-parameters. This paper introduces a general framework for interactively exploring the similarities and differences among a set of dependency networks and demonstrates our work-in-progress on a specific implementation for multiple Bayesian networks.