TADC-SBM: a Time-varying, Attributed, Degree-Corrected Stochastic Block Model

We present a synthetic dataset generator that produces temporal graphs with varying community structures, attribute features, and temporal dynamics, allowing for the evaluation of node clustering methods in a systematic manner. Temporal graphs offer a robust framework for modeling dynamic systems, with far-reaching applications in various domains where the analysis of evolving relationships between entities over time is required, such as transportation networks and recommendation systems. However, detecting communities in such graphs poses significant challenges, as the underlying community structure is subject to change over time and the presence of additional node or edge attributes introduces further complexity. Recent advances in graph neural networks have shown promise for ”neural” community detection, but their expressiveness and generalization capabilities in attributed temporal graphs remain unclear, largely due to the scarcity of suitable real-world datasets for evaluation. In an experimental evaluation using TADC-SBM, we observe that novel approaches for node clustering can display good performance in scenarios with low community stability, but do not consistently outperform most baselines, highlighting potential research opportunities and underscoring the need for more generalizable models and robust benchmarks and datasets.