Towards More Reliable Machine Learning: Conceptual Insights and Practical Approaches for Unsupervised Manifold Learning and Supervised Benchmark Studies
MCML Authors
Moritz Herrmann
Dr.
Transfer Coordinator
Abstract
Moritz Herrmann
Dr.
Transfer Coordinator
Abstract
This thesis focuses on improving the reliability and trustworthiness of machine learning, particularly in unsupervised learning methods like manifold learning. It investigates the challenges of evaluating manifold learning techniques and proposes improvements for embedding evaluation, outlier detection, and cluster analysis, using methods like UMAP and DBSCAN. Additionally, the thesis contributes to supervised learning by presenting a benchmark study on survival prediction in multi-omics cancer data and exploring the effects of design and analysis choices on benchmark results. (Shortened).
BibTeXKey: Her22