HIVE: A Hyperbolic Interactive Visualization Explorer for Representation Learning

Thijmen Nijdam, Derck W. E. Prinzhorn, Jurgen de Heus, Thomas Brouwer

Keywords: Hyperbolic Embeddings, Multimedia Analytics, Interactive Visualization, Visual Analytics for AI

TL;DR: HIVE is an interactive dashboard for exploring hyperbolic embeddings, supporting multimodal analysis and comparison of different projection methods.

Abstract: We present HIVE, an interactive dashboard that supports exploration and interpretation of hyperbolic embeddings in deep learning. Hyperbolic spaces naturally capture hierarchical structure, yet existing visualization tools are either designed for Euclidean geometry or remain static when curvature is taken into account. HIVE closes this gap by offering 2D projections in the Poincaré disk and integrating configurable dimensionality-reduction algorithms for hyperbolic space, including CO-SNE and HoroPCA. From expert interviews, we distilled key analytic needs and realized them in four interaction modes: compare, traverse, tree, and neighbors. These modes enable real-time, multimodal analysis through semantic hierarchy tracing, geodesic interpolation, and projection comparison. A small but targeted user study demonstrates that HIVE supports practical analysis and uncovers meaningful hyperbolic structure. While currently limited to image and text embeddings, the dashboard shows promise for broader applications, such as reinforcement learning and graph discovery, highlighting HIVE's potential as a useful tool for future hyperbolic learning scenarios. Source code and a demo are available at: https://github.com/thijmennijdam/HIVE.

Track: Full paper (8 pages excluding references, same as main conference requirements)

Git: https://github.com/thijmennijdam/HIVE

Submission Number: 8