Decision-making processes increasingly rely on complex machine learning models, but their predictive power often comes at the expense of transparency, limiting our ability to understand and learn from predictions. Such models achieve their accuracy by exploiting multivariate feature relationships, yet these relationships remain difficult to interpret: high-dimensional spaces prevent direct inspection, dimensionality reduction can distort neighborhoods and decision boundaries, and existing tools rarely connect model behavior to domain knowledge. This dissertation addresses this gap by contributing three model-agnostic visual analytics techniques for multivariate tabular data, organized around complementary perspectives of model interpretation. Input-based analysis: Decision boundaries, i.e., regions where a model’s prediction flips under multivariate input changes, are hard to explore in high dimensions. This dissertation presents an interactive system that systematically probes input perturbations using local linear maps, preserving input distances more accurately and revealing closer boundaries than prior approaches. Relationship-based analysis: Visual feature enrichment in non-linear dimensionality reductions struggles to balance multiple objectives, such as representing clusters, outliers and distortion. This dissertation introduces a topology-based augmentation method that effectively relates feature distributions to projection regions, simultaneously highlighting clusters, outliers, and ambiguities. Knowledge-based analysis: Matrix completion yields prediction matrices whose patterns are challenging to interpret beyond basic heatmap expansion and sorting. This dissertation proposes a hierarchical evaluation framework that aggregates and links these patterns to domain-knowledge not used during model building. Its application yielded insights that substantially advanced thermodynamic modeling. All approaches are integrated in comprehensive visual analytics systems designed for interactive use. Case studies, stakeholder evaluations, and successful applications confirm their effectiveness in promoting the explainability of model predictions and domain insight. Together, these contributions establish structured methods for exploring multivariate feature spaces visually, driving measurable advances in both model and data interpretability.
Jan‐Tobias Sohns (Thu,) studied this question.