IEEE VIS 2026 paper maps network roles, challenges industry label assumptions

A short paper accepted at IEEE VIS 2026 (arXiv:2607.02163) tackles a classic data mining challenge: exploring structural equivalence—nodes that occupy similar roles in a network—in attributed networks. Authors Kohei Arimoto and Masahiko Itoh argue that popular representation learning methods implicitly assume a uniform, continuous feature space. Instead, they propose a visual analytics approach using dimensionality reduction to reveal the true topological structure of high-dimensional feature spaces derived from nodes' neighborhood attribute profiles (i.e., the set of features describing a node's local connections).

Applying their method to inter-firm transaction networks, they discovered that structural roles cluster into complex, non-linear manifolds with notable density biases. Comparing these feature space manifolds against conventional industry classifications yielded three key insights: (1) supply chain hierarchies transition smoothly rather than discretely, (2) categories that share general semantic labels separate clearly when viewed through actual transaction patterns, and (3) a single industry label can fragment into multiple distinct regions. These results challenge the assumption that identical semantics imply similar structural roles and highlight the need for new similarity metrics aligned with manifold topology, not just Euclidean distance. The work is only 5 pages with 3 figures, setting the stage for richer follow-up studies.