TL;DR
Fedivertex is a new open-source dataset comprising 182 graph snapshots across seven Fediverse platforms (including Mastodon, Lemmy, and PeerTube), tracked weekly over 14 weeks. It captures server-to-server interactions rather than user-level feeds, allowing researchers to study decentralized machine learning and a unique real-world phenomenon: defederation (edge deletion).
What Problem Does This Paper Address?
Decentralized machine learning (P2P training where nodes collaborate without a central server) is growing rapidly to enhance data privacy and remove single points of failure. However, an algorithm's convergence speed, privacy guarantees, and robustness depend heavily on the underlying communication topology.
Existing benchmarks rely almost exclusively on centralized, for-profit networks (like Twitter or Facebook) or synthetic topologies. These traditional datasets suffer from key flaws:
- They are static snapshots, failing to capture temporal shifts.
- They operate at the individual user level, where link formation is heavily skewed by algorithmic recommendation engines.
- They assume networks only grow via new connections (triadic closure), ignoring intentional disconnection.
The Fediverse—an open, federated ecosystem run by independent server administrators using the ActivityPub protocol—offers an ideal testbed. However, prior Fediverse datasets have been fragmented, single-platform, or non-reusable.
Methods
The authors built a web crawler operating on top of the curated Fediverse Observer directory. To capture the ecosystem's diversity, they constructed several graph types:
- Federation graphs: Undirected graphs where an edge exists if two instances have interacted at least once.
- Active user graphs: Weighted, directed graphs tracking the following relationships of each instance's 10,000 most active users.
- Intra- and cross-instance interaction graphs: Directed and undirected interaction networks derived from comment threads (e.g., on Lemmy).
To protect user privacy, the collection is strictly instance-level, avoids storing post text, respects robots.txt and #NOBOT flags, and rate-limits requests to 0.4 seconds per instance. The dataset is bundled with a PyPI package (fedivertex) and provides exports directly compatible with NetworkX and Gephi.
Using these graphs, the authors benchmarked three tasks:
- Decentralized Differentially Private Learning: Simulating private random-walk SGD across instances.
- Defederation Prediction: A newly formalized task of predicting which edges or nodes will be severed in subsequent snapshots.
- Community Detection: Evaluating clustering algorithms against ground-truth instance languages.
Datasets
- Fedivertex (Introduced): 182 graphs spanning 13 different graph formulations across 7 platforms (PeerTube, Mastodon, Pleroma, Misskey, Friendica, Bookwyrm, Lemmy), crawled weekly across 14 timestamps over three months.
- Comparison Datasets (from SNAP): Wikipedia Vote, Facebook Ego, Twitter Ego, and GitHub collaboration graphs.
Key Insights and Ideas
- Non-Power-Law Topologies: While commercial networks exhibit power-law degree distributions (rich-get-richer dynamics), Fedivertex graphs feature concave degree curves with shorter tails, reflecting a more balanced distribution of node influence.
- Severing Ties as a Network Dynamic: Unlike mainstream social networks where edge deletion is negligible, Fediverse instances actively "defederate" over moderation disputes, technical failures, or security concerns. Some graphs (such as Pleroma federation or Misskey active users) were dominated by edge loss rather than growth during the study period.
- Inverted Heuristics Fail on Deletions: Predicting edge deletion by reversing standard link-prediction heuristics (e.g., picking lowest common neighbors or Jaccard scores) performs poorly on federation graphs, demonstrating that cutting ties is driven by factors beyond simple structural dissimilarity.
- Convergence Bottlenecks in Decentralized ML: On Mastodon's active-user graph, random-walk DP-SGD converges significantly slower than on synthetic regular graphs of identical size (3,852 nodes), largely because poorly connected peripheral nodes are rarely visited.
Key Results
- Link Addition vs. Deletion Prediction: Top-$K$ topological heuristics easily outperform random baselines for edge additions (e.g., on Misskey federation Top-1000, Adamic-Adar scores 494 vs. $42 \pm 6$ for random). However, for edge deletions, topological heuristics barely match or slightly beat random guessing (Adamic-Adar scores 62 vs. $51 \pm 3.7$ for random).
- Decentralized DP-SGD: Pairwise privacy loss between instances closely mirrors graph communicability and Katz centrality, confirming theoretical predictions on real topologies.
- Community Detection: Using primary instance language (top 5 classes):
- On PeerTube Follow, Louvain achieved an Adjusted Mutual Information (AMI) of 0.123 and Modularity of 0.2168.
- On Misskey Active User, Label Propagation reached an Adjusted Rand Index (ARI) of 0.229 and AMI of 0.255, while Louvain delivered the highest Modularity (0.611). No single algorithm dominated across all metrics.
Limitations and Future Work
- Temporal Granularity: A weekly crawl interval is too coarse to capture rapid defederation cascades that unfold in hours or days.
- Purely Structural Baseline: Defederation relies heavily on moderation and social context; topological metrics alone provide insufficient predictive power for edge deletion.
- Future Directions: The authors plan to continue longitudinal crawling, increase sampling frequency during defederation spikes, and expand applications into early detection of instance shutdowns and decentralized spam filtering.
Summary updated