Deep community detection in attributed temporal graphs: experimental evaluation of current approaches

Recent advances in network representation learning have sparked renewed interest in developing strategies for learning on spatio-temporal signals, crucial for applications like traffic forecasting, recommendation systems, and social network analysis. Despite the popularity of Graph Neural Networks for node-level clustering, most specialized solutions are evaluated in transductive learning settings, where the entire graph is available during training, leaving a significant gap in understanding their performance in inductive learning settings. This work presents an experimental evaluation of community detection approaches on temporal graphs, comparing traditional methods with deep learning models geared toward node-level clustering. We assess their performance on six real-world datasets, focused on a transductive setting and extending to an inductive setting for one dataset. Our results show that deep learning models for graphs do not consistently outperform more established methods on this task, highlighting the need for more effective approaches and comprehensive benchmarks for their evaluation.