Methods for Transparency and Control in Embedding Visualizations

Abstract

Embedding visualizations compress complex high-dimensional data into two dimensions and have become standard practice for exploring high-dimensional data. Yet they are easy to misread, since the relationship between the projection and the underlying data, and the meaning of the structures it surfaces, is often unclear. We develop three approaches that improve transparency at different stages of the embedding visualization generation. First, the Hypertrix computes and visualizes geometric distortions introduced by dimensionality reduction. Second, the Explain-and-Test framework uses machine learning to generate and verify explanations for user-specified clusters. Third, Concept-guided embedding manipulation enables modification of the embedding space to produce alternative organizations of the same data. Together, these contributions add geometric faithfulness, semantic transparency, and organizational control to embedding visualizations.