Abstract

Visualization and machine learning research are both driven by a desire to extract insights from data. However, the means to this end differ substantially between the two fields. While machine learning typically tries to automate decisions, visualization focuses on the human in the loop. A combination of these disparate approaches can help users to acquire insights from data more effectively. This thesis compiles results from five studies in which visualization and machine learning were brought together with a focus on temporal and/or high-dimensional processes. These works span the range from visualizations for model analysis to data processing for visualization. ConfusionFlow and InstanceFlow are two interactive visualization systems that let machine learning developers analyze the temporal progression of the training of classification models. A more general analysis of high-dimensional, temporal processes is possible with the Projection Path Explorer, which visualizes processes as trajectories in a low-dimensional embedding space. The Projection Path Explorer makes use of unsupervised machine learning for nonlinear dimensionality reduction. Projective Latent Interventions show how these unsupervised techniques can be adapted to give users more control over, and a better understanding of, the latent representations of classification models. To this end, parametric extensions of dimensionality reduction techniques are introduced, which allow users to manipulate the embeddings in such a way that changes can be propagated back to the original classification model. Finally, ParaDime is a framework for specifying such parametric dimensionality reduction routines in a flexible and customizable way. ParaDime unifies existing techniques and facilitates experimentation with new embedding methods for visualization. These five works illustrate the variety of possible combinations of machine learning and visualization, and showcase how such combined approaches can help users to better understand high-dimensional, temporal processes.

Citation

BibTeX

@thesis{2022_phd_thesis_hinterreiter,
    title = {Visual Explanations of High-dimensional and Temporal Processes},
    author = {Andreas Hinterreiter},
    institution = {Johannes Kepler University Linz},
    month = {November},
    type = {phdthesis},
    year = {2022}
}

Acknowledgements

This PhD was carried out within the Human-Interpretable Machine Learning project, a collaboration between Johannes Kepler University Linz and Imperial College London funded by the State of Upper Austria. Additional financial support by Johannes Kepler University Linz, the Linz Institute of Technology (LIT), the State of Upper Austria, and the Federal Ministry of Education, Science and Research (LIT-2019-7-SEE-117), by the Austrian Science Fund (FWF P27975-NBL and FWF DFH 23–N), by the Austrian Research Promotion Agency (FFG 851460 and FFG 854184), by the Wellcome Trust (IEH 102431 and EPSRC EP/S013687/1.), and by the Boehringer Ingelheim Regional Center Vienna is gratefully acknowledged.