Interactive Exploration of Multitask Dependency Networks

Scientists increasingly depend on machine learning algorithms to discover patterns in complex data. Two examples addressed in this dissertation are identifying how information sharing among regions of the brain develops due to learning; and, learning dependency networks of blood proteins associated with cancer. Dependency networks, or graphical models, are learned from the observed data in order to make comparisons between the sub-populations of the dataset. Rarely is there su cient data to infer robust individual networks for each sub-population. The multiple networks must be considered simultaneously; exploding the hypothesis space of the learning problem. Exploring this complex solution space requires input from the domain scientist to refine the objective function. This dissertation introduces a framework to incorporate domain knowledge in transfer learning to facilitate the exploration of solutions. The framework is a generalization of existing algorithms for multiple network structure identification. Solutions produced with human input narrow down the variance of solutions to those that answer questions of interest to domain scientists. Patterns, such as identifying di↵erences between networks, are learned with higher confidence using transfer learning than through the standard method of bootstrapping. Transfer learning may be the