Static-Graph Estimation in Hybrid GNNs Using Behavioral-Similarity and Clustering-Based Methods
IEEE Access · 2026 · 14 · 42734-42750
DOI: 10.1109/access.2026.3674781 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Ryusei Otani und 4 weiteren Autor:innen. Die Arbeit nennt 34 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Japan Society for the Promotion of Science (JSPS) KAKENHI, Japan Science and Technology Agency (JST) Core Research for Evolutionary Science and Technology, Outstanding Research Group Support Program of Nagoya City University.
Abstract (Originalsprache)
Hybrid GNNs, which learn both long-term structural information encoded in static graphs and temporal interactions within dynamic graphs, have attracted attention for their high predictive accuracy. In practical applications, however, the static relational data required as part of the input is not always available. Furthermore, many existing static-graph construction methods rely on domain-specific prior knowledge or node attributes; the former hinders generalization to other domains, while the latter is difficult to apply when attribute information is scarce. To address these issues, this paper proposes a framework that enables the application of hybrid GNNs by estimating a static graph as a preprocessing step using only dynamic graphs as input. We introduce three static-graph construction methods: two co-occurrence-based methods and one clustering-based method. The co-occurrence-based methods construct static graphs based on “behavioral similarity” derived from dynamic interactions. Specifically, they aggregate node co-occurrence frequencies over a certain period and connect nodes with similar behavioral tendencies. The clustering-based method applies K-means clustering to node representations derived from dynamic interactions and constructs a sparse static graph within each cluster using the k-nearest neighbor method. These approaches enable the construction of effective static graphs for hybrid GNNs even when attribute information is limited. Evaluations using five real-world datasets demonstrated that the proposed methods improve the predictive accuracy and training efficiency of hybrid GNNs compared to attribute-based construction methods. The results further highlight the importance of selecting an appropriate static-graph construction method based on data characteristics and application scenarios.
Förderung
- Japan Society for the Promotion of Science (JSPS) KAKENHI
- Japan Science and Technology Agency (JST) Core Research for Evolutionary Science and Technology
- Outstanding Research Group Support Program of Nagoya City University
Zitieren
APA
Otani, R., Namikoshi, K., Sakurai, Y., Guo, M., & Oyama, S. (2026). Static-Graph Estimation in Hybrid GNNs Using Behavioral-Similarity and Clustering-Based Methods. IEEE Access, 14, 42734-42750. https://doi.org/10.1109/access.2026.3674781
Vancouver
Otani R, Namikoshi K, Sakurai Y, Guo M, Oyama S. Static-Graph Estimation in Hybrid GNNs Using Behavioral-Similarity and Clustering-Based Methods. IEEE Access. 2026;14:42734-42750. doi:10.1109/access.2026.3674781
BibTeX
@article{Otani2026,
author = {Otani, Ryusei and Namikoshi, Keiichi and Sakurai, Yuko and Guo, Mingyu and Oyama, Satoshi},
title = {Static-Graph Estimation in Hybrid GNNs Using Behavioral-Similarity and Clustering-Based Methods},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {42734--42750},
doi = {10.1109/access.2026.3674781}
}
EUVIMED