CRII: III: Integrating Domain Knowledge via Interactive Multi-Task Learning

Michigan State University, East Lansing MI

Investigators

Abstract

The ever increasing availability of data has attracted a huge amount of effort building machine learning models from the data to unleash its hidden power. One ubiquitous finding about these machine learning tasks is that in most real-world applications the learning tasks are closely related to each other. Moreover, human experts in many domains can usually provide indispensable domain knowledge describing how these models are related. Maximally exploiting such knowledge is critical in building high quality machine learning models. This project will develop effective and efficient interactive algorithms and tools (including open source software) to enable knowledge discovery by integrating domain knowledge of task relatedness from human experts. The algorithms and tools developed in this project will directly impact biomedical informatics as they will be used to build disease progression models. The educational component of this project includes developing a new curriculum that incorporates research into the classroom and provides students from under-represented groups with opportunities to participate in research. Leveraging task relatedness, multi-task learning (MTL) simultaneously learns all related learning tasks and performs knowledge transfer among the tasks to improve the quality of models from all the tasks. Although there are numerous studies for MTL that assume different types of task relatedness, limited progress has been made in incorporating domain knowledge in MTL. This project will advance MTL by: (1) developing algorithms for knowledge aware multi-task feature learning which exploit domain knowledge of features to guide the selection of joint features from the learning tasks; (2) developing algorithms for knowledge aware multi-task relationship learning which utilize domain knowledge of tasks to guide the learning of task relationships; and (3) developing efficient and scalable optimization algorithms to facilitate effective interactive visualization. For further information see the project web page: http://jiayuzhou.github.io/projects/crii