Explaining and Steering Embedding Projections for Visual Analytics

Abstract

Low-dimensional embedding projections are widely used in visual analytics, particularly for exploring large document collections. By arranging documents as points in a two-dimensional space, these projections help analysts identify clusters, outliers, separations, and relationships among documents. However, projection layouts are often difficult to interpret and control: users can observe where documents are positioned, but may not understand why spatial patterns appear or how to reshape the projection when the resulting layout does not align with their analytic goals. This dissertation frames embedding projections as interactive semantic workspaces and develops methods for explaining and steering them in visual analytics. First, it introduces gradient-based explanations that connect textual features to document positions in projection layouts, revealing how words influence spatial placement. Second, it presents context-aware natural-language explanations that combine document semantics with layout-derived spatial context to help users interpret documents, regions, and spatial patterns. Third, it moves from explanation to steering by introducing an LLM-augmented semantic steering approach, in which analysts express semantic intent through example groupings and reshape projections without retraining the underlying models. Finally, it develops a scalable prototype-based steering method that shifts LLM reasoning from individual items to group-level abstraction, making semantic steering practical for large embedding collections. Through quantitative evaluations, usage scenarios, case studies, and a user study, this dissertation demonstrates that embedding projections can be made more interpretable, controllable, and aligned with analytic goals. These contributions advance projection-based visual analytics toward interactive semantic workspaces that analysts can inspect, understand, and reshape.