Do Servers Matter on Mastodon?

Data-driven Design for Recommendations in Decentralized Social Media

Carl Colglazier

Community Data Science Collective

Northwestern University

@carl@hci.social

June 3, 2024

Mastodon’s growth highlights opportunities and challenges in the decentralized web

Figure 1: Accounts in the dataset created between January 2022 and March 2023. The top panels shows the proportion of accounts still active 45 days after creation, the proportion of accounts that have moved, and the proportion of accounts that have been suspended. The bottom panel shows the count of accounts created each week. The dashed vertical lines in the bottom panel represent the announcement day of the Elon Musk Twitter acquisition, the acquisition closing day, a day where Twitter suspended a number of prominent journalist, and a day when Twitter experienced an outage and started rate limiting accounts.

Here come the newcomers

Caveat: how do we determine success (for servers) in a decentralized social network?

Avoiding the “X killer hype cycle” where

  1. A writer discovers an alternative technology system

  2. Media hypes it as a “killer” of a major platform

  3. The system does not in fact “kill” the major platform

  4. The system is declared a failure

This has happened mutliple times already (Zulli, Liu, and Gehl 2020).

Mastodon does not need to replace something else to be successful

We should instead take social communities on their own terms.

Do people find value in the system?

Decentralization means location matters more

Each server has its own:

Working assumptions:

We thus ask: do some servers retain newcomers better than others?

Smaller, less general servers are more likely to retain new accounts

Applying a survival model to accounts created in May 2023, we find accounts on the 12 largest servers featured on joinmastodon.org are more likely to become inactive in the first 91 days; further, we find accounts on smaller servers are less likely to become inactive.

Figure 2: Survival probabilities for accounts created during May 2023.
Term Estimate 95% CI p-value
Join Mastodon 0.115 (0.97, 1.3) 0.117
General Servers 0.385 (1.07, 2.01) 0.017
Small Server -0.245 (0.66, 0.92) 0.003

Cox Proportional Hazard Model with Mixed Effects. The model includes a random effect for the server.

Accounts that move between servers are more likely to move to smaller servers

Table 1
Model A
Model B
Coef. Std.Error Coef. Std.Error
Sum -9.529 ***0.188 -10.268 ***0.718
Nonzero -3.577 ***0.083 -2.861 ***0.254
Smaller server 0.709 ***0.032 0.629 ***0.082
Server size (outgoing) 0.686 ***0.013 0.655 ***0.042
Open registrations (incoming) 0.168 ***0.046 -0.250 0.186
Languages match 0.044 0.065 0.589 0.392

Exponential family random graph models for account movement between Mastodon servers. Accounts in Model A were created in May 2022 and moved to another account at some later point. Accounts in Model B were created at some earlier point and moved after October 2023.

Our analysis suggests…

  • Accounts on large, general servers fare worse
  • Moved accounts go to smaller servers

Can we build a system that helps people find servers?

Constraints

  • Consent: servers should be able to choose whether to participate

  • Privacy: do not reveal information about individual accounts

  • Decentralization: do not concentrate data in one place

  • Openness: use shared standards and protocols

Concept

A decentralized, tag-based collaborative filtering system

  • Each server reports their top tags from the last three months

  • Learn from these reports and from other servers which tags are most important for each server

  • Recommend servers based on selected tags of interest

Implementation

  • Report top hashtags used by the most accounts on each server

  • For robustness, drop hashtags used by too few accounts or servers

  • Build an \(m \times n\) server-tag matrix \(M\)

  • Normalize with Okai BM25 TF-IDF and L2 normalization

  • Apply singular value decomposition (SVD) on \(M\) to create a new matrix \(M'\)

  • Match servers to selected tags using cosine similarity

Demo

https://carlcolglazier.com/demos/deweb2024/

Thanks!

Special thanks to my qualifying exam committee: Noshir Contractor, Darren Gergle, and Aaron Shaw; to the Community Data Science Collective; and to the Technology and Social Behavior program at Northwestern University for funding this research.

References

Colglazier, Carl, Nathan TeBlunthuis, and Aaron Shaw. 2024. “The Effects of Group Sanctions on Participation and Toxicity: Quasi-experimental Evidence from the Fediverse.” arXiv. https://doi.org/10.48550/arXiv.2404.02109.
Gehl, Robert W., and Diana Zulli. 2023. “The Digital Covenant: Non-Centralized Platform Governance on the Mastodon Social Network.” Information, Communication & Society 26 (16): 3275–91. https://doi.org/10.1080/1369118X.2022.2147400.
Kraut, Robert E., Paul Resnick, and Sara Kiesler. 2011. Building Successful Online Communities: Evidence-Based Social Design. Cambridge, Mass: MIT Press.
Nicholson, Matthew N., Brian C Keegan, and Casey Fiesler. 2023. “Mastodon Rules: Characterizing Formal Rules on Popular Mastodon Instances.” In Companion Publication of the 2023 Conference on Computer Supported Cooperative Work and Social Computing, 86–90. CSCW ’23 Companion. New York, NY, USA: Association for Computing Machinery. https://doi.org/10.1145/3584931.3606970.
Zulli, Diana, Miao Liu, and Robert Gehl. 2020. “Rethinking the ‘Social’ in ‘Social Media’: Insights into Topology, Abstraction, and Scale on the Mastodon Social Network.” New Media & Society 22 (7): 1188–1205. https://doi.org/10.1177/1461444820912533.

Evaluation

How well does the recommender work?

A brief history of Mastodon

timeline
  title Mastodon and Fediverse Timeline
  2008: OStatus Protocol
  2016: Mastodon releases v0.1
  2018: ActivityPub standard published
  2019: Mastodon drops OStatus
  2022: Elon Musk Twitter acquisition
      : Truth Social launches using Mastodon code
  2023: Mastodon reaches 2M active users 
      : Threads (Meta) begins experimental support for ActivityPub