Craig Newgard
Oregon Health & Science University, OR·Computer Science
Prediction models depend on features, the measured characteristics that describe something. For a patient, these might be age, smoking status or years of smoking.
- Fields
Computer Science
Engineering
- Topics
feature generation
feature engineering
automated machine learning
- Grants
- 1 known grant, through Aug 2027
What they work on
- Develop a reinforcement learning framework in which two agents select features and a third selects an operation, with a shared value network, to generate new features in a traceable way.
- Design heuristic selection strategies using information-theory-based utility scores to evaluate features and feature sets, making automated feature generation computationally affordable.
- Fine-tune a large language model on tabular data converted to natural-language text, with compressed prompts, so it generates new features together with human-readable explanations.
More about the research
Combining features can make predictions better. For example, age minus years smoked shows how early someone started smoking. Building such combinations by hand (feature generation) is not practical when datasets contain huge numbers of features. This project, part of a collaborative grant titled around closing the "sim-to-real" gap in reinforcement learning, aims to generate new features automatically. It also aims to make the process traceable, affordable in computing cost, and explainable.
Why it matters
Automated, explainable feature generation could reduce reliance on human experts when building prediction models from large datasets. It may improve prediction accuracy in areas such as healthcare risk prediction, intelligent transportation, social analysis and internet-of-things systems. Explanations of the generated features could make these models easier to trust and audit.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Chi Jin
Princeton University, NJ·Computer Science·Early-career
Identify the general structures of function classes, such as those represented by neural networks, under which model-free reinforcement learning is tractable, and design provably efficient algorithms that also handle model misspecification.
Alvaro Velasquez
University of Colorado Boulder, CO·Computer Science
Develop decentralized self-organization methods in which AI agents decompose tasks among peers, assign and enhance roles, and self-refine during execution, using information-theoretic tests to judge whether emergent behavior is useful.
Haipeng Chen
William & Mary, VA·Computer Science
Develop physics-inspired, task-aware graph neural networks that build problem structure into reinforcement learning, so it can handle badly shaped or implicit objectives in combinatorial optimization.
The school
Oregon Health & Science University: 24 labs on PI Pages, $545M a year on research, 54 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.