Contributions in Interpretable and Trustworthy Machine Learning
Name
liu-briliu-phd-orc-2026-thesis.pdf
Size
7.49 MB
Format
Adobe PDF
Checksum (MD5)
5be55700c47b36c050a7b6d4580fdc12
Author(s)
Liu, Brian
Advisor(s)
Mazumder, Rahul
Date Issued
May 2026
Publisher
Massachusetts Institute of Technology
Abstract
Modern statistical and machine learning models are recognized for their predictive accuracy and ability to capture complex interactions. However, while these models excel at prediction, they are often difficult to interpret and may fail to reveal useful relationships in the data. In this thesis, we explore how discrete optimization can improve interpretability and trustworthiness in statistical modeling. In Chapters 2 and 3, using tree ensembles (e.g., gradient boosting and random forests) as an example, we propose an optimization-based framework for extracting interpretable models (rule sets) at the post-training stage. Our estimator is defined as the solution to a large-scale discrete optimization problem that balances multiple considerations, including predictive accuracy, compactness, and stability. We develop specialized algorithms to solve these problems at a scale beyond the capabilities of off-the-shelf optimization software. In Chapter 4, we propose a new optimization-based framework for fitting piecewise constant additive models. Our framework scales to problem sizes beyond the reach of other popular approaches and can explicitly accommodate structural constraints such as feature sparsity. Finally, in Chapter 5, we explore the role of feature subset randomization in random forests. We demonstrate that this randomization can simultaneously reduce both bias and variance in random forests, offering an alternative perspective to the prevailing view that random forests primarily act by regularizing bagging ensembles.
MIT Department
Massachusetts Institute of Technology. Operations Research Center
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
