In many areas, such as citation networks and co-purchase graphs, analyzing complex networks requires understanding both graph structure and node attributes. Traditional dimensionality reduction methods, such as t-SNE, handle only one data source at a time, limiting the ability to capture the full information in networks. Structural and attribute information are complementary, and combining them can significantly improve understanding of node relationships and cluster structures. To address this limitation, we propose a modification of t-SNE that can handle both data sources simultaneously. State-of-the-art graph neural networks already embed both types of information, but we propose an additional step in the DR process. We also introduce a novel approach to help to explain why nodes appear close or distant in the visualization through an interactive tool. These approaches lead to cleaner cluster separation, more interpretable patterns, and high harmonic trustworthiness, which is our main evaluation metric.
Pedro Gagini (Wed,) studied this question.