Representation Learning and Temporal Modeling of Large-Scale Dynamic SystemsPoličar, Pavlin Gregor (Avtor) Zupan, Blaž (Mentor) Žitnik, Marinka (Komentor) dimensionality reductiont-SNEinterpretabilitytemporal modelingelectronic health recordssingle-cell genomicsModern scientific inquiry is increasingly characterized by the analysis of massive, high- dimensional datasets. In many domains, low-dimensional embeddings have become an important tool for exploring such data: they reveal structure, support hypothesis generation, and facilitate scientific communication. However, their practical use raises methodological challenges related to scalability, reuse, interpretation, and the incorporation of temporal structure. This dissertation makes three sets of contributions. First, it develops scalable and reusable embedding infrastructure through openTSNE, a modular Python implementation of t-SNE, and a principled method for embedding new observations into existing reference maps, enabling comparative analyses and helping mitigate batch effects in single-cell genomics. Second, it introduces automated methods for interpreting embeddings, including VERA for static region-based explanations and a domain-aware annotation approach that incorporates hierarchical medical knowledge. Third, it develops methods for temporally aware embedding construction, including the directional coherence loss and related extensions that incorporate temporal structure directly into the embedding process. Evaluated on nationwide electronic health record and single-cell transcriptomics data, these contributions show how low-dimensional embeddings can be constructed, reused, interpreted, and extended to support reasoning about complex dynamic systems.20262026-08-06 14:35:01Doktorsko delo/naloga185497sl