Abstract
Decentralized social media platforms like Bluesky Social (Bluesky) have made it possible to publicly disclose user behaviors with millisecond-level precision. Embracing Bluesky's principles of open-source and open-data, we present BlueTempNet, the first collection of the temporal dynamics of user-driven social interactions on the platform. BlueTempNet integrates multiple types of networks into a single multi-network, including user-to-user interactions (following and blocking) and user-to-community interactions (creating and joining custom Feeds). This dataset captures past user behaviors and provides a framework for future longitudinal studies of decentralized social ecosystems.
Introduction
Studying human behavior on social media has traditionally been hampered by two "black box" problems. First, platform-driven recommendation algorithms—such as the "For You" page on X or Facebook—manipulate user behavior, making it nearly impossible to isolate organic social dynamics. Second, major platforms severely restrict API access, often hiding "negative" interactions like blocking or providing coarse, non-temporal snapshots of follower lists.
Bluesky Social, built on the decentralized AT Protocol (ATproto), offers a rare "glass box" environment. Because the platform is designed for data transparency and user-driven moderation, it allows researchers to observe how users form and dissolve connections without the interference of opaque corporate algorithms. BlueTempNet leverages this architecture to provide a comprehensive, longitudinal dataset of social interactions.
Data Collection and Multi-network Architecture
The researchers developed a three-tiered data collection pipeline using the ATproto API to capture the platform's multi-layered social structure:
- Community Level: Identification of active "Feeds"—user-created custom algorithms that function as decentralized communities.
- User-to-Community Level: Data on "Feed Creators" (those who build the algorithms) and "Feed Members" (those who "Like" or subscribe to them).
- User-to-User Level: Specific relationships between users, including both positive (follows) and negative (blocks) interactions.
By treating the data as a multi-network, the researchers integrated these different interaction types into a single framework. In this model, nodes (users) can have multiple types of edges (e.g., User A follows User B while also joining Community C). All data points include millisecond-precision timestamps, enabling high-resolution temporal analysis of network evolution.
Dataset Scale and Scope
The BlueTempNet dataset covers a full year of activity, from May 11, 2023, to May 11, 2024. Its scale and granularity distinguish it from existing datasets derived from platforms like Reddit or Mastodon:
- Scale: The dataset includes over 150,000 public users and nearly 40,000 custom Feeds.
- Edge Types: It contains approximately 5.1 million "Follow" edges and 460,000 "Block" edges.
- Temporal Precision: Unlike static snapshots, BlueTempNet offers "perfect coverage" of interaction history with exact timestamps, allowing for the study of real-time network shifts.
Key Insights into Social Dynamics
The Visibility of Blocking
One of the most significant contributions of BlueTempNet is the ability to study blocking behavior, a metric usually hidden by centralized platforms. The data reveals that while follows are more common in absolute terms, blocking grew at a faster rate than following after Bluesky opened to the public in February 2024. This suggests that as a platform scales, "negative" social maintenance becomes a primary activity for users managing their digital environment.
Feed Economics and Community Engagement
There is a sharp contrast between what users create and what they consume. While niche feeds (such as those dedicated to "Art" or specific "Media") are the most numerous, feeds centered on "Discovery" and "Trending" topics receive the vast majority of likes and engagement. This indicates a power-law distribution in community attention despite the decentralized nature of feed creation.
Network Resilience and Small-World Properties
User-to-user interactions (follows and blocks) proved more resilient and stable over time than user-to-community interactions. When the platform shifted from invitation-only to public, community-joining behavior showed massive variability, while the underlying social graph grew more steadily. Furthermore, by projecting community interactions into a "line graph," the researchers discovered that Bluesky’s custom feeds exhibit "small-world" properties—high clustering and short path lengths—particularly around language-based and interest-based clusters, such as Japanese art or Ukrainian political discussion.
Structural Analysis Results
- Assortativity: The network exhibits slightly negative degree assortativity. This implies that high-profile users (influencers) tend to be followed by lower-degree users rather than exclusively by other influencers.
- Structural Balance: The "Feed Member" network shows a higher proportion of "balanced triangles" than the "Feed Creator" network. This suggests that regular users are more likely to achieve social equilibrium in their circles than the developers of the feeds themselves.
- Growth Spikes: Following the platform's public launch on February 6, 2024, the number of follows in the "Creator" network jumped by 19.2%, while blocks surged by 27.3%.
Limitations and Future Directions
As a relatively young platform, Bluesky's user base is currently dominated by "migrants" from X, which may bias the data toward specific niche interests like art and technology. This "platform inertia" means the behaviors observed may not yet represent the "average" social media user.
Future research using BlueTempNet will focus on sentiment analysis of the text and rumors spreading through these networks, as well as causal modeling to determine if specific blocking events trigger wider shifts in network stability. Finally, cross-platform studies will compare how the same users behave on Bluesky versus Mastodon or X to determine how decentralized architecture truly alters human interaction.
Summary updated