Abstract: Fedivertex is a new open-source dataset featuring 182 temporal graphs from seven decentralized social networks (the "Fediverse"), designed to benchmark decentralized machine learning (DML) and graph analysis. Beyond standard tasks, it introduces "defederation prediction"—the challenge of predicting when nodes or edges will be removed from a network—providing a more realistic and ethically aligned alternative to datasets from centralized, for-profit platforms.
Introduction
Decentralized machine learning (DML) allows multiple participants to train models collaboratively without a central server, keeping data local and improving scalability. However, the success of DML is heavily dependent on the topology of the communication graph. Most existing graph datasets are either synthetic or crawled from centralized platforms like Facebook or Twitter. These "for-profit" graphs are problematic because their links are heavily influenced by proprietary recommendation algorithms, which don't reflect the organic, trust-based collaborations DML aims to model. Furthermore, most datasets are static snapshots that ignore the temporal dynamics of how networks grow and, crucially, how they shrink.
Methodology
The authors developed a custom web crawler to collect data from seven major Fediverse platforms: Mastodon, Misskey, Peertube, Pleroma, Friendica, Bookwyrm, and Lemmy. Weekly snapshots were taken over 14 weeks to capture temporal evolution.
Key methodological contributions include:
- Diverse Graph Construction: The authors built different types of graphs, including "federation graphs" (server-to-server connections) and "active user graphs" (based on follow relationships).
- Defederation Formalization: They formally defined the task of predicting node and edge deletions, a process often driven by moderation disagreements or technical failures in decentralized systems.
- Software Tooling: The release of the
fedivertexPython package integrates with NetworkX to make the data easily accessible for researchers.
The Fedivertex Dataset
The primary contribution is the Fedivertex dataset, which consists of 182 graphs total (13 distinct graph types, each with 14 weekly snapshots). It covers seven platforms spanning video streaming (Peertube), microblogging (Mastodon, Misskey, Pleroma, Friendica), book cataloging (Bookwyrm), and social news (Lemmy). Nodes include metadata like the official language of the instance, which serves as a ground truth for community detection tasks.
Structural Insights
Analysis reveals that Fedivertex graphs exhibit different growth dynamics than traditional social networks. While centralized platforms often follow a strict power-law decay in degree distribution, Fedivertex graphs show more diverse "concave" attachment dynamics. This suggests that decentralized networks grow differently than algorithmically-driven ones.
Further insights include:
- Topology and Convergence: DML algorithms converge significantly slower on real-world Fediverse graphs than on synthetic ones, likely due to "low-centrality" nodes on the periphery.
- Language as a Proxy for Community: In the Fediverse, language is a primary driver of network structure. For example, Misskey is dominated by Japanese-speaking communities, while Peertube has a strong French-speaking cluster.
- The "Defederation" Challenge: Link deletion is a major dynamic. Instances often "defederate" (sever ties) due to moderation conflicts. Predicting this is much harder than predicting new links; standard heuristics like Jaccard similarity often perform no better than random guessing for deletions.
Key Results
Fedivertex graphs exhibit "small-world" properties (high clustering, low path length), but with a much wider range of average path lengths (1.18 to 3.95) compared to SNAP datasets. Using a private decentralized logistic regression task, the authors showed that the specific topology of a Mastodon graph can drastically slow down convergence compared to a synthetic graph of the same size. Community detection algorithms like Louvain and Label Propagation were tested using language as ground truth, showing varying performance across platforms and indicating diverse structural complexities.
Conclusion
Fedivertex addresses the need for realistic, decentralized graph data. While the current weekly temporal granularity may miss rapid events, it provides a robust baseline for future research. The authors plan to continue updating the dataset and explore more sophisticated models for defederation prediction that incorporate node metadata or higher-order network properties.
Summary updated