Accelerating scientific research: Empowering automated knowledge discovery through machine learning

Automatic knowledge discovery is vital for the progress of scientific research. For example, tackling missing data is especially needed in health care and social network domains, and being able to model complex chemical structures at the text level brings enormous benefits to chemistry and biomedical domains. We aim to develop machine learning-based frameworks that are capable of generating novel knowledge by learning from massive real-world data to help accelerate scientific research. The first part of the thesis introduces GATE, a graph-based variational auto-encoder framework, developed in response to the prevalent issue of missing node features in networks and that traditional methods have fallen short in adequately addressing this challenge. The second part of the thesis tackles the intricate task of predicting fine-grained chemical entity types from chemical literature, a critical component in advancing biomedical and chemical research. It presents a novel multi-modal representation learning framework and a newly created benchmark dataset, CHEMET. This framework leveraged external resources with chemical structures and used cross-modal attention to learn an effective representation of text in the chemistry domain. Experiments show that our approaches significantly outperform state-of-the-art methods.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2023

Author dc:creator

• Sun, Chenkai

Contributors dc:contributor

• Ji, Heng
• Zhai, ChengXiang

• Machine Learning

Statement dc:rights

• Copyright 2023 Chenkai Sun

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/122252

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F122252