Fedivertex: a Graph Dataset based on Decentralized Social Networks for Trustworthy Machine Learning

Marc Damie, Edwige Cyffers · 2025

Read Fedivertex: a Graph Dataset based on Decentralized Social Networks for Trustworthy Machine Learning by Marc Damie, Edwige Cyffers (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-20882.txt.

Introduction

Marc Damie and Edwige Cyffers release Fedivertex, a dataset of network graphs drawn from decentralized social media. Their 2025 paper aims to give machine learning researchers real graphs for testing learning methods where users keep their own data. Older graph datasets mostly come from big commercial platforms frozen at one moment.

Fedivertex holds 182 graphs from seven networks in the Fediverse, the family that includes Mastodon, Misskey, and Lemmy. The data was gathered weekly over fourteen weeks, so change over time is visible. A Python package ships with the data. The authors show uses such as defederation, where one community cuts links with another, a pattern unique to these networks.

Catalog record

CREATORMarc Damie, Edwige Cyffers
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE51 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-20882

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Fedivertex: a Graph Dataset based on Decentralized Social Networks for Trustworthy Machine Learning Marc Damie, Edwige Cyffers Abstract Decentralized machine learning – where each client keeps its own data locally and uses its own computational resources to collaboratively train a model by exchanging peer-to-peer messages – is increasingly popular, as it enables better scalability and control over the data. A major challenge in this setting is that learning dynamics depend on the topology of the communication graph, which motivates the use of real graph datasets for benchmarking decentralized algorithms. Unfortunately, existing graph datasets are largely limited to for-profit social networks crawled at a fixed point in time and often collected at the user scale, where links are heavily influenced by the platform and its recommendation algorithms. The Fediverse, which includes several free and open-source decentralized social media platforms such as Mastodon, Misskey, and Lemmy, offers an interesting real-world alternative. We introduce Fedivertex , a new dataset of 182 graphs, covering seven social networks from the Fediverse, crawled weekly over 14 weeks. We release the dataset along with a Python package to facilitate its use, and illustrate its utility on several tasks, including a new defederation task, which captures a process of link deletion observed on these networks. 1 Introduction Decentralized machine learning hallaji_decentralized_2024 has gained significant popularity in the past years. In this paradigm, each node possesses a local dataset and some computational resources, and collaborates with other participants by exchanging messages through peer-to-peer communication during the training of a global, potentially personalized, machine learning model. In comparison to federated learning kairouz_advances_2021 , which keeps data local but orchestrates training through a central server, decentralized learning offers additional flexibility as it avoids the bottlenecks and single points of failure that arise from centralized supervision. The shift towards decentralized learning can also be motivated by trust, as the communication graph can reflect users’ chosen collaborations–often referred to as nodes, instances, clients, or participants in this context. The network topology has an impact on the learning dynamics kavalionak_impact_2021 , particularly in the presence of data heterogeneity hsieh_non-iid_2020 ; vogels_relaysum_2021 and in terms of privacy guarantees mrini2024privacy . Decentralized learning has numerous real-world use cases hallaji_decentralized_2024 , as nodes can represent healthcare institutions or sensors distributed across installations. One of the most compelling applications is for decentralized social networks bin_zia_toxicity_2022 . In such cases, the graph often captures complex and diverse relationships, which explains its popularity in machine learning. In particular, this motivates various graph learning tasks tang_graph_2010 , including community detection, node classification, and edge prediction. Social network dynamics also raise interesting questions related to polarization and time-evolving properties of the graph toivonen_comparative_2009 . Addressing these questions requires access to relevant social network datasets that allow studying these properties. The Fediverse – a contraction of “federation” and “universe” – provides decentralized and interoperable online social services. It is often seen as an alternative anderlini_emerging_2022 to major social networks operated by for-profit companies, and it promotes a very different culture. The Fediverse is decentralized across many servers, called instances . Anyone can run an instance, which operates independently under the moderation of its owner, and instances collaborate with each other: a user of a given instance can interact with and follow users from other instances. Since 2018, the Fediverse has adopted ActivityPub activitypub , a protocol and open standard that provides a client-to-server API for creating and modifying content, as well as a federated server-to-server protocol for delivering notifications and content across servers. This enables interoperability between different instances and software. The diversity of platforms, the growing number of users, and the international impact of the Fediverse make it an interesting object of study for the machine learning community. In particular, agents in the Fediverse tend to be more aware of the potential ethical issues of machine learning than traditional users of social networks Struett2023 , and more interested in new features and improvements. This aligns closely with the goals of Trustworthy Machine Learning and the paradigm of collaborative learning, where agents are expected to monitor their participation based on expected benefits. In this work, we provide the first dataset covering multiple software platforms in the Fediverse, called Fedivertex , to enable researchers to easily run experiments on decentralized machine learning tasks and to benchmark several graph learning tasks. By surveying seven different platforms and constructing different types of graphs, we are able to capture the diverse dynamics at play in the Fediverse. In particular, a striking difference from many mainstream social networks is the so-called defederation process lai_new_2025 , in which instances choose to sever ties with other instances, often due to a disagreement on moderation or security practices. While new link prediction is often regarded as the primary task for time-evolving graphs hasan_survey_2011 , a major dynamic in the Fediverse is this complementary edge deletion. Our dataset is the first to enable the study of this phenomenon. The current 14 distinct timestamps for each graph are a starting point to study the evolution over time, and we plan to continue to update the dataset in the future. More precisely, our contributions are as follows: (i) We introduce Fedivertex , a large and diverse graph dataset based on the Fediverse. More precisely, our dataset encompasses seven Fediverse platforms, resulting in 182 graphs: 13 different graphs each with a sequence of 14 snapshots obtained through weekly web crawls over a period of three months. (ii) We provide a Python package, fedivertex , available through PyPI, to easily access and use our dataset. The package includes built-in preprocessing tools to download and prepare the graphs for machine learning tasks. We demonstrate its usefulness by benchmarking several existing decentralized learning algorithms. (iii) We formalize a novel graph analysis task: defederation prediction , which aims to predict which edges or nodes will be removed from the graph at the next iteration, and we propose baselines for this task. 2 Related work Decentralized machine learning. Federated learning and fully decentralized learning are increasingly studied ( kairouz_advances_2021 ; Li2020 ; pmlr-v54-mcmahan17a ) , with various algorithms based on gossip ( boyd2006gossip ; hendrikx_principled_2022 ; Hendrikx2020DualFreeSD ; pmlr-v119-koloskova20a ; dec2017 ; JMLR:v17:15-292 ; neglia2020 ; tang18a ) or random walks ( even_stochastic_2023 ; johansson ; hendrikx_principled_2022 ; Walkman ) . These results highlight the importance of communication graphs for the quality of the final model, the speed of convergence in the presence of heterogeneous data ( pmlr-v206-le-bars23a ) , personalization Bellet2018a and privacy guarantees ( cyffers2024differentially ; cyffers2022muffliato ; Dekker2023TopologyBasedRP ; mrini2024privacy ) , which motivates the use of recent real-world social networks. Social network datasets and analysis. Machine learning frequently relies on small social networks, such as the Karate Club ( karate ) or citation networks ( citeseer ; cora ) . Several larger digital social networks are also available via platforms like SNAP ( leskovec_snap_2014 ) , in particular Facebook and Twitter graphs. It has been shown that for-profit platforms influence user graphs, as their recommendations about whom to follow accelerate the triadic closure process and exacerbate inequality in popularity ( effectofrecommendation ; zignani_follow_2018 ) . This motivates the study of social networks that do not follow this trend. In particular, the Fediverse enables analysis at the

Read the complete text