Evaluating Representation Learning and Graph Layout Methods for Visualization.

Area of Science:

• Machine Learning
• Data Visualization
• Graph Theory

Background:

• Representation learning methods embed graph nodes into low-dimensional spaces, enhancing machine learning tasks.
• These learned representations are often assumed suitable for graph visualization, but lack empirical validation.
• Existing research lacks comprehensive benchmarks for evaluating graph representation learning in visualization contexts.

Purpose of the Study:

• To empirically compare state-of-the-art representation learning methods against graph layout algorithms for graph visualization.
• To assess visualization quality using readability, distance-based metrics, and link prediction performance.
• To identify the most effective methods for generating informative and interpretable graph visualizations.

Main Methods:

• Comparison of multiple representation learning techniques with two recent graph layout algorithms.
• Evaluation using quantitative measures: readability scores and distance-based metrics.
• Assessment of visualization effectiveness via link prediction performance on embedded graphs.

Main Results:

• No single representation learning method consistently outperformed others across all quality measures.
• Graph layout algorithms generally produced qualitatively superior layouts compared to representation learning methods.
• Higher dimensional embeddings with t-distributed stochastic neighbor embedding improved local neighborhood preservation but increased computational cost.

Conclusions:

• Graph layout algorithms are currently more effective for graph visualization than representation learning methods.
• The assumption that low-dimensional embeddings are optimal for visualization requires further investigation.
• Optimizing visualization quality involves trade-offs between neighborhood preservation and computational efficiency.