Project Overview

The ever-increasing availability of data has led to a proliferation of machine learning models built to unlock its hidden potential. A ubiquitous finding across real-world applications is that learning tasks are closely related to one another. Moreover, human experts in many domains can often provide indispensable insights — yet traditional machine learning pipelines rarely incorporate this knowledge in a structured, interactive way.

This NSF CRII project addresses that gap by developing frameworks for interactive multi-task learning, where domain knowledge from human experts is integrated directly into the model-building process. The research aims to produce algorithms and open-source software tools that make multi-task learning more accessible and effective, with a particular focus on biomedical informatics applications such as disease progression modeling.

Multi-task learning has emerged as a powerful paradigm in modern machine learning, particularly suited to domains where multiple related prediction problems must be solved simultaneously. Rather than building isolated models for each individual task, this approach leverages shared structure across tasks to improve generalization, especially when data is limited for certain tasks. In biomedical settings, for example, different but related diseases may share underlying biological mechanisms, and multi-task models can exploit these commonalities to produce more robust and accurate predictions. This methodology has proven valuable in analyzing electronic health records, where numerous clinical outcomes and patient subgroups are naturally interconnected and can benefit from joint modeling.

Interactive machine learning represents a growing shift toward systems that incorporate human expertise directly into the model-building process. Traditional pipelines often treat data as static and domain knowledge as external, but interactive frameworks allow experts to guide feature selection, label refinement, and model structure iteratively. In medical informatics, clinicians can provide insights about disease progression patterns or relevant biomarkers that may not be evident from raw data alone. By closing the loop between human judgment and algorithmic learning, these approaches yield models that are not only more accurate but also more interpretable and aligned with real-world clinical reasoning and decision-making.

Large-scale data analytics in biomedical informatics presents unique challenges that demand efficient and scalable machine learning solutions. From high-dimensional genomic data to longitudinal electronic health records, the volume and complexity of modern medical data require algorithms that can handle millions of features and thousands of tasks simultaneously. Sparse learning methods, feature selection techniques, and factorization models have been developed to extract meaningful patterns from such data while maintaining computational feasibility. These advances have enabled breakthroughs in understanding disease progression, identifying patient subtypes, and building predictive models for conditions such as Alzheimer's disease, ultimately supporting more personalized and data-driven approaches to healthcare.

Research & Educational Goals

The project pursues several interconnected objectives:

• Developing novel algorithms that enable human experts to interactively inject domain knowledge into multi-task learning frameworks.
• Building disease progression models that leverage integrated domain expertise for biomedical informatics.
• Producing open-source software tools to lower the barrier for researchers and practitioners adopting multi-task learning methods.
• Incorporating educational components to train students in large-scale machine learning and its real-world applications.