Visual Analytics for Temporal Hypergraph Model Exploration.

Area of Science:

• Visual Analytics
• Machine Learning
• Network Science

Background:

• Complex systems (e.g., gene interactions, social networks) are better modeled by temporal hypergraphs than traditional graphs.
• Hypergraphs generalize graphs by allowing edges to connect multiple vertices, enabling more accurate representation of intricate relationships.
• Interactive exploration and refinement of hypergraph-based prediction models remain a significant challenge.

Purpose of the Study:

• To introduce Hyper-Matrix, a novel visual analytics technique for exploring and refining temporal hypergraph models.
• To address the challenge of interactive exploration and seamless refinement of complex hypergraph-based prediction models.
• To tightly couple machine learning with interactive visualizations for enhanced user-driven exploration.

Main Methods:

• Developed Hyper-Matrix, a technique integrating geometric deep learning with interactive visualizations.
• Employed a matrix-based visualization with multi-level semantic zoom for scalable drill-down capabilities.
• Incorporated user-steering interactions for filtering, search, dynamic partition hierarchy, reordering, and model feedback.

Main Results:

• Hyper-Matrix demonstrates superior scalability and applicability compared to existing solutions.
• The technique facilitates the incorporation of domain knowledge and enables rapid search-space traversal.
• Evaluated through a case study and formative evaluation with law enforcement experts using real-world data.

Conclusions:

• Hyper-Matrix provides an effective solution for the visual analytics of temporal hypergraphs.
• The approach enables user-driven exploration and refinement of complex hypergraph models.
• Paves the way for applying visual analytics to temporal hypergraphs across various domains.