Designing trustworthy machine learning systems with structure‑aware conformal inference Wenwen Si

Format:

Book

Thesis/Dissertation

Author/Creator:

Si, Wenwen, author.

Contributor:

University of Pennsylvania. Computer and Information Science., degree granting institution.

Language:

English

Subjects (All):

Computer science.

Bioinformatics.

Information science.

0984.

0800.

0723.

0715.

Local Subjects:

Computer science.

Bioinformatics.

Information science.

0984.

0800.

0723.

0715.

Genre:

Academic theses

Physical Description:

1 online resource (148 pages)

Contained In:

Dissertations Abstracts International 87-12A

Place of Publication:

Ann Arbor : ProQuest Dissertations and Theses, 2026

Language Note:

English

Summary:

The safe deployment of machine learning systems requires reliable uncertainty quantification. Conformal prediction has emerged as a promising framework because it provides attractive finite-sample reliability guarantees. However, classical conformal methods are often inadequate for real-world deployment, where data distributions may shift, prediction targets are often structured, and decisions are made sequentially over time. In this dissertation, we study how to design trustworthy machine learning systems by combining conformal inference with problem-specific structure across three major settings: (1) prediction under label shift, where target label proportions are unknown, (2) structured medical forecasting, where uncertainty must be clinically interpretable and robust under domain shift, and (3) sequential decision-making, where reliability and efficiency must be balanced over multiple steps. First, we develop a method for constructing training-set conditional, probably approximately correct prediction sets under label shift that accounts for confidence intervals around importance weights by modifying Gaussian elimination to propagate intervals. Second, for longitudinal glaucoma visual field progression prediction, we introduce archetype-based conformal inference, which represents uncertainty through clinically meaningful progression patterns, and we further extend this framework to domain shift through shift-aware conformal risk training. Third, we develop conformal approaches for reliable sequential decision-making, including conformal constrained policy optimization for language-model orchestration, policy-guided stepwise reasoning, and uncertainty-aware conformal training of a continuous-time decision transformer for sequential glaucoma medication prediction on the UPenn dataset. Across these settings, the dissertation shows how conformal inference can be aligned with distributional, archetypal, and sequential structure to produce outputs that are not only statistically valid, but also informative, interpretable, and useful for downstream decision-making

Notes:

Source: Dissertations Abstracts International, Volume: 87-12, Section: A.

Advisors: Lee, Insup; Al-Aswad, Lama Committee members: Bastani, Osbert; Alur, Rajeev; Wong, Eric; Weimer, James

Ph.D. University of Pennsylvania 2026

Vendor supplied data

Local Notes:

School code: 0175

ISBN:

9798247973614

Access Restriction:

Restricted for use by site license