Understanding the evolution of online communities is crucial for analyzing social dynamics on platforms during politically sensitive periods. Using Louvain, HDBSCAN, and Ward clustering, we analyze community structures on a private X dataset collected during an election period, and employ LLM-based labeling to assess semantic coherence, finding that Ward’s clusters are the most interpretable. In addition, we propose TopoTemp, a novel framework combining topological data analysis with sequential modeling to predict community evolution in temporal social networks, achieving consistent improvements over baselines. Furthermore, we experiment on both the X and Reddit Hyperlinks dataset to determine the impact of snapshot-level, as well as community-level, topological features when applied to community evolution prediction. This work highlights the importance of data granularity and feature selection in dynamic network analysis and offers a practical foundation for future research in community evolution prediction.
the cluster labels we can clearly see that HDBSCAN has quite detailed and inquisitive labels like "Breaking News", "Canada National News", and "Sarcasm". While the other clustering methods are broader in nature with labels like "Political Figures", "Religious Group", and "LGBTQ+ Advocacy". Overall, Ward produces more intuitive and well-separated clusters, often capturing clear distinctions such as liberal versus conservative politics. In contrast, Louvain tends to identify broader, less specific groupings, typically labeling clusters with high-level themes like political leader, group, or figure, without distinguishing between ideological alignments. Furthermore, among the three algorithms, Louvain resulted in the highest number of "unknown" clusters, while Ward’s has the fewest.
