Ye, Ye
(2019)
Transfer Learning for Bayesian Case Detection Systems.
Doctoral Dissertation, University of Pittsburgh.
(Unpublished)
Abstract
In this age of big biomedical data, a variety of data has been produced worldwide. If we could combine that data more effectively, we might well develop a deeper understanding of biomedical problems and their solutions. Compared to traditional machine learning techniques, transfer learning techniques explicitly model differences among origins of data to provide a smooth transfer of knowledge. Most techniques focus on the transfer of data, while more recent techniques have begun to explore the possibility of transfer of models. Model-transfer techniques are especially appealing in biomedicine because they involve fewer privacy risks. Unfortunately, most model-transfer techniques are unable to handle heterogeneous scenarios where models differ in the features they contain, which occur commonly with biomedical data. This dissertation develops an innovative transfer learning framework to share both data and models under a variety of conditions, while allowing the inclusion of features that are unique to and informative about the target context. I used both synthetic and real-world datasets to test two hypotheses: 1) a transfer learning model that is learned using source knowledge and target data performs classification in the target context better than a target model that is learned solely from target data; 2) a transfer learning model performs classification in the target context better than a source model. I conducted a comprehensive analysis to investigate conditions where these two hypotheses hold, and more generally the factors that affect the effectiveness of transfer learning, providing empirical opinions about when and what to share. My research enables knowledge sharing under heterogeneous scenarios and provides an approach for understanding transfer learning performance in terms of differences of features, distributions, and sample sizes between source and target. The model-transfer algorithm can be viewed as a new Bayesian network learning algorithm with a flexible representation of prior knowledge. In concrete terms, this work shows the potential for transfer learning to assist in the rapid development of a case detection system for an emergent unknown disease. More generally, to my knowledge, this research is the first investigation of model-based transfer learning in biomedicine under heterogeneous scenarios.
Share
| Citation/Export: |
|
| Social Networking: |
|
Details
| Item Type: |
University of Pittsburgh ETD
|
| Status: |
Unpublished |
| Creators/Authors: |
|
| ETD Committee: |
|
| Date: |
8 January 2019 |
| Date Type: |
Publication |
| Defense Date: |
12 November 2018 |
| Approval Date: |
8 January 2019 |
| Submission Date: |
5 December 2018 |
| Access Restriction: |
No restriction; Release the ETD for access worldwide immediately. |
| Number of Pages: |
167 |
| Institution: |
University of Pittsburgh |
| Schools and Programs: |
School of Computing and Information > Intelligent Systems Program |
| Degree: |
PhD - Doctor of Philosophy |
| Thesis Type: |
Doctoral Dissertation |
| Refereed: |
Yes |
| Uncontrolled Keywords: |
transfer learning, Bayesian, case detection, heterogeneous, model transfer, influenza |
| Date Deposited: |
08 Jan 2019 19:38 |
| Last Modified: |
08 Jan 2019 19:38 |
| URI: |
http://d-scholarship.pitt.edu/id/eprint/35684 |
Metrics
Monthly Views for the past 3 years
Plum Analytics
Actions (login required)
 |
View Item |
![[feed]](/images/feed-icon-32x32.png){kind=icon}