Graph Generation via Temporal-Aware Biased Walks

Some real networks keep a fixed structure (e.g., roads, sensors and their connections) while node or edge signals evolve over time. Existing graph generators either model topology changes (i.e., edge additions/deletions) or focus only on static graph properties (such as degree distributions or motifs), without considering how temporal signals shape the generated structure. By approaching the problem from an unconventional perspective, we introduce TANGEM, that integrate a temporal similarity matrix into biased random walks, thereby coupling signals with structure to generate graphs that highlight patterns reflecting how nodes co-activate over time. We evaluate TANGEM using an approach that separates structural fidelity (clustering, spectral metrics) from downstream temporal consistency, allowing us to clearly isolate the impact of the topology generator itself. In structural benchmarks, TANGEM consistently outperforms strong baselines while remaining lightweight. These results show that adding attribute-guided bias to structural sampling produces more realistic graphs and establishes TANGEM as a basis for future models that further integrate evolving signals and structure.