Efficient Neural Community Detection on Time-Varying Attributed Graphs

Community detection or node clustering is a fundamental task in network analysis, where mesoscale structure may help reveal the function and dynamics of a complex system. Many and diverse applications - including social network analysis, computational biology, and recommendation systems - rely on community detection for goals such as node classification, link prediction, and graph summarization. Motivated by the need for more robust and scalable methods across domains, this research focuses on the task in time-varying attributed graphs - where the network's topology may be subject to change, and nodes and edges may be associated with additional metadata. Its contributions are three-fold: (i) software infrastructure, datasets, and generators for modeling and analyzing temporal graphs; (ii) high-performance algorithms for temporal community detection, based on spectral clustering or heuristic optimization on GPU devices; and (iii) a novel machine learning model that incorporates both spatio-temporal and attribute signals in its objective, improving upon a state-of-the-art graph neural network for end-to-end differentiable community detection. In doing so, this thesis intersects the network science and computer science domains in order to answer its central research question: can neural strategies outperform established baselines? Systematic evaluation on both real-world and synthetic datasets demonstrates that the advantages of neural methods over algorithmic approaches are contingent rather than universal: significant improvements are observed only where temporal or attribute signals carry information that a static, aggregated view of the network would otherwise miss, and provided those signals align with the community structure being recovered. Where this condition does not hold, established algorithms may remain competitive and outperform neural baselines, especially in regimes well studied in the literature; there, their decisive limitation lies not in accuracy, but in tractability and scalability to large temporal graphs or those with a high temporal resolution. These findings position learning-based and classical methods as complementary, the better choice depending on the properties of the dataset at hand - reinforcing a no-free-lunch view of community detection, in which no single method is universally optimal and the most suitable one varies with the data. Together, these contributions advance community detection in temporal graphs by introducing new methods, investigating the conditions under which neural or algorithmic methods are most effective, and providing new tools and resources for further research in the area.