Year: 2021
Type: dissertation
Abstract: In this thesis we explore a wide range of statistical learning algorithms and evaluate their abilities toanswer clear and precise research questions given a real data set. We do so in multiple research fields; we tackle a higher education problem, contribute to oncology research and analyse an image data set. Our evaluation of algorithms is made with respect to their performance on these real data sets and their overall accessibility. This creates opportunities for interesting findings and results and it serves as motivation for theoretical and algorithmic contributions. We revisit the theoretical foundation of algorithms in order to adapt them to tackle well-established statistical problems. We also visit every step of data analysis; from data collection passing by algorithm design to a thorough analysis that answers research questions. This gives us a complete perspective of the entire data analysis pipeline which allows us to form constructive criticisms about algorithms. In this thesis we discuss all of our recent contributions. We analyse a higher education data set and demonstrate the ability of algorithms to accurately predict students in needs of support. Inspired by the structure of this data set, we construct a new decision algorithm and provide a R-package to anyone interested in using this algorithm. We study the gap between the theory and the implementation of variational autoencoders, illustrate these differences on real data sets and provide explanations and recommendations on how to fix these issues. We design a new survival analysis model using the variational autoencoder framework and evaluate multiple survival analysis machine learning models on the largest randomized trial in pediatric Hodgkin Lymphoma ever conducted. We collect and share a new data set containing high-resolution images of hand-written digits attached to a collection of covariates. Finally, we demonstrate the potential of controllable generative models; a new perspective on the use of labelled data sets and generative models to provide user control over the generated variables.
Source: TSpace (University of Toronto)
Author: Beaulac, Cédric
Institution: University of Toronto
Language: English
Subfield: Statistics and Probability
Field: Mathematics
Domain: Physical Sciences
SDG: Quality education
Open Access status: green
OpenAlex ID: W7133060953
