NNP-NET: Accelerating t-SNE Graph Drawing for Large Static and Dynamic Graphs by Neural Networks.

Area of Science:

• Computer Science
• Data Visualization
• Machine Learning

Background:

• Recent graph drawing (GD) methods like tsNET produce high-quality layouts but are computationally expensive due to reliance on t-SNE.
• There is a need for efficient graph drawing algorithms that can handle large-scale and dynamic graph data without sacrificing layout quality.

Purpose of the Study:

• To introduce NNP-NET, a novel graph drawing method that addresses the runtime limitations of tsNET.
• To adapt the NNP projection technique for efficient and high-quality layout generation of both static and dynamic graphs.

Main Methods:

• NNP-NET adapts the NNP (Neighbor-based Non-linear Projection) technique for graph projection, enabling linear scaling with data size.
• The method handles both unweighted and weighted graphs and leverages NNP's out-of-sample capability for dynamic graph projection.
• Layout quality is optimized to be comparable to tsNET while significantly improving computational efficiency.

Main Results:

• NNP-NET demonstrates significantly faster performance compared to existing methods for very large graphs (up to 50 million nodes and 108 million edges).
• The projected layouts achieve quality metrics close to the ground-truth tsNET.
• For dynamic graphs, NNP-NET effectively balances layout stability with high visual quality.

Conclusions:

• NNP-NET provides an efficient and effective solution for drawing large-scale and dynamic graphs.
• The method offers a compelling alternative to t-SNE-based approaches, delivering comparable quality at a fraction of the computational cost.
• NNP-NET advances the field of graph drawing by enabling the visualization of complex, time-varying network structures.