PaperlayerView exact revision on arXiv

Source-linked AI summary

Out-of-Network Attention Dynamics on Bluesky

Andrea Failla, Veronica Mesina, Giulio Rossetti

arXiv:2610.01971v1cs.SIcs.CY

TL;DR

How much attention crosses beyond users’ explicit follow ties, and whether such attention becomes new ties, remains understudied at scale. Using Bluesky interaction records, the paper measures network distance, exploratory behavior, and follow conversion; most interaction stays near the existing graph, and distant exploration converts less often.

  • Problem

    How far individual acts of attention extend beyond users’ follow neighborhoods remains understudied, despite the importance of distinguishing network structure from content curation.

    Source: §1 · p. 2

  • Method

    The study analyzes observed interactions by attention route and follow distance, clusters users’ exploration behavior and trajectories, and models follow conversion for explored user–author pairs.

    Source: §3 · p. 4

  • Results

    80.8% of scored interactions land within two follow hops versus 22.4% under the expected-degree null, while distant contacts convert at 3.3% versus 11.9% for local exploratory contacts.

    Source: §5 · p. 9

  • Takeaways & Limitations

    On Bluesky, exploration is substantial in volume but mostly stays near the existing graph, and its capacity to reshape the follow graph appears limited.

    Source: §5 · p. 9

    Interaction records capture acted-upon attention rather than exposure, locality is reconstructed from the available follow graph, conversion models are observational, and the study covers Bluesky’s early growth period.

    Source: §5 · p. 9

Abstract

from arXiv · show

Personalized social media commonly relies on explicit follow graphs to shape what content users encounter; yet how much attention crosses ties they have not formed remains largely undocumented at scale. We study this question on Bluesky, a large decentralized microblogging platform whose default feed relies on a simple, reverse-chronological content recommender. We analyze 173 million user-author interactions (likes, reposts, replies, and quotes) collected from a near-complete platform dump between February and September 2023. We decompose each interaction by attention-path length (already followed, relayed by a followed account, reachable within two follow-hops, or beyond) and find that 74.5% of interactions reach the user through an account they already follow. Measured by distance in the follow graph rather than by route, 80.6% of interaction lands within two follow hops, far beyond the 22.4% an expected-degree null predicts. We then characterize how exploration varies across users and over tenure. A broad-reaching minority generates three quarters of all exploratory activity, while aggregate declines in exploration with tenure mask three distinct individual trajectories. Finally, attention reaching beyond two hops converts into new follow ties at less than one third the rate of two-hop-local exploratory attention. Together, these results depict a platform where out-of-network exploration is substantial in volume but strongly constrained by network proximity and unlikely to translate into new social ties.

1 Introduction

The paper uses Bluesky’s reverse-chronological Following feed as a meaningful baseline for studying attention beyond users’ explicit ties. It asks how far attention reaches, how exploration varies, and whether it becomes new ties.

1 Introduction

  • Bluesky’s default Following feed is reverse chronological, making its follow graph a meaningful baseline for studying attention beyond users’ explicit ties.Other recommendation mechanisms are user-selectable alternatives, with uptake limited relative to the platform’s user base.Source: §1 · pp. 1–2
  • The study asks how much attention escapes the follow graph, how exploration varies across users and tenure, and whether escaped attention converts into new ties.The introduction frames this gap against selective social ties and algorithmic curation, which can concentrate exposure around existing interests and relationships.Source: §1 · p. 2
  • 74.5% of interactions arrive through followed authors or followed-account reshares, and 80.6% land within two follow hops, above the expected-degree null.The study analyzes 173 million interactions by 1.1 million users.Source: §1 · p. 2

2 Related Work

Prior work links online attention to network structure, selective exposure, and tie formation, but has rarely measured how far individual attention travels from a user’s follow neighborhood. This study centers that structural distance on Bluesky.

Network structure and attention

  • Prior work links homophily, selective exposure, and confirmation bias to homogeneous neighborhoods, while studies often measure concentration at community, ideological, or diffusion levels rather than individual follow distance.The cited work describes clustered and sometimes ideologically segregated interaction patterns.Source: §2 · p. 3

Boundary-spanning interaction and tie formation

  • Weak and bridging ties may connect distant network regions and facilitate novel information, but boundary-spanning contact does not necessarily become a persistent tie.Prior work on evolving networks links tie formation to network position, interaction, and triadic closure.Source: §2 · p. 3

Bluesky

  • Earlier Bluesky studies describe platform growth, network structure, participation, and mechanisms such as Starter Packs; this paper instead examines interaction’s spatial relation to the follow graph.The cited studies include work on activity and follower networks, topology and algorithmic feeds [15], Starter Packs [5], and participation styles [12].Source: §2 · p. 3

3 Methods

The study uses Bluesky activity records to measure observed interaction by network route and distance, characterize exploratory behavior, and assess whether explored user–author pairs become follow ties.

Data Collection

  • The analysis uses the first eight months of Bluesky activity, February through September 2023, drawn from available records queried from sampled users’ Personal Data Servers.The initial sample was approximately 5 million users; activity records were queried via the Bluesky API.Source: §3 · pp. 3–4

Attention-Path Length

  • Observed attention is measured through likes, reposts, replies, and quotes; these interactions indicate attended content, but unacted exposure is unobserved, making reported quantities lower bounds.Interactions are classified as in-network, relay-mediated, two-hop local exploratory, or distant exploratory; relay identifies a possible route, not actual delivery.Source: §3 · p. 4

Clustering

  • Two clusterings characterize users’ exploration composition and their own eight-week exploration-rate trajectories, with trajectory rates z-scored to remove overall level.The trajectory clustering distinguishes rising, falling, or transient behavior.Source: §3 · p. 4

Conversion Rate Analysis

  • Conversion analysis measures first-follow outcomes at 7, 15, and 30 days and uses adjusted pair-level logistic regression for 30-day conversion without claiming causality.Pairs are split by whether first contact was two-hop local or distant and by first-event type.Source: §3 · p. 4

4 Results

Most interactions route through followed accounts, while attention reaches two-hop authors far more often than an expected-degree null predicts. Exploratory activity is concentrated among broad-reaching users, and distant first contacts convert less often than local ones.

4.1 RQ1: How Much Attention Escapes the Follow Graph?

  • 74.5% of interactions route through followed authors or relays, while 80.6% land within two follow hops versus 22.4% under the expected-degree null.The two-hop distance measure excludes relays; the observed share exceeds the null by 58.4 points.Source: Figure 1 · p. 5, §4.1 · p. 5

4.2 RQ2: How is Exploration Distributed Across Users, and Does it Change with Tenure?

  • 22.7% of users in the broad, low-repeat cluster generated 74.4% of explored events; the highest-rate cluster reached only 31.6 authors on average.The highest-rate cluster comprised 18.8% of users, with 66.7% of activity exploratory.Source: §4.2 · p. 6, Figure 2 · p. 7
  • The pooled exploration rate fell from 24.3% to 21.0%, but trajectories split into downward (34.6%), upward (35.0%), and transient (30.4%) clusters.The downward cluster moved from 38.1% to 20.7%, the upward cluster from 16.3% to 25.0%, and the transient cluster peaked at 26.7% in week 2.Source: §4.2 · p. 6, Figure 3 · p. 7

4.3 RQ3: Does Escaped Attention Convert Into New Ties?

  • 11.9% of two-hop-local first-contact pairs converted within 30 days versus 3.3% of distant pairs; adjusted distant-contact odds were 0.351 relative to local (95% CI 0.335–0.368).The adjusted model accounts for first-event type and seven user-level covariates.Source: §4.3 · pp. 7–8, Table 1 · p. 8, Table 2 · p. 8

5 Discussion and Conclusion

On Bluesky, most interaction remains close to users’ existing follow networks, while distant exploration is less likely to become a new tie. Exploration also varies substantially across users, and the study’s observational and early-platform setting limits how broadly to interpret these patterns.

5 Discussion and Conclusion

  • 80.6% of all interactions land within two follow hops; among scored interactions, the share is 80.8% versus 22.4% under the expected-degree null.The study describes network proximity as a powerful correlate of where interaction occurs.Source: §5 · pp. 8–9
  • Distant contacts convert into follows at 3.3%, compared with 11.9% for local exploratory contacts, and the difference remains after adjustment for observed user characteristics.The authors conclude that distant attention is less likely to enter users’ subsequent social networks.Source: §5 · p. 9
  • A relatively small group generates most exploratory activity, while declining, increasing, and transient exploration trajectories coexist across users.The aggregate decline in exploration with tenure therefore conceals individual heterogeneity.Source: §5 · p. 9
  • The authors distinguish exploration as exposure from structural change: exploration is substantial, but its capacity to reshape the follow graph appears limited.Encounters traveling further from users’ existing networks rarely produce new ties.Source: §5 · p. 9
  • Interaction records capture acted-upon attention rather than exposure, two-hop locality is reconstructed from the available graph, and observational models do not establish causality.The study covers Bluesky’s early growth, so these patterns may change as the platform and recommendation ecosystem evolve.Source: §5 · p. 9

A Robustness, Validation, and Model Specifications

The analyses test the distance–tie formation association against alternative nulls and possible confounds, while documenting the clustering and conversion-model choices. They also specify limits in the model’s uncertainty estimates.

A Robustness, Validation, and Model Specifications

  • Across early, mid, and late cohorts, within-two-hop shares (82.0%, 78.8%, 80.3%) exceeded expected-degree nulls (28.2%, 32.5%, 28.0%).The null is evaluated at each scored interaction’s timestamp; its comparison excludes relay-mediated routes because relays are information routes, not follow paths.Source: A Robustness, Validation, and Model… · pp. 9–10
  • Composition clustering was stable under resampling (ARI 0.984), while trajectory clustering used K = 3 for stronger cross-model agreement (ARI 0.509 vs. 0.226).The trajectory solution remained similarly reproducible under resampling (split-half ARI 0.956 vs. 0.968 for K = 2).Source: A Robustness, Validation, and Model… · p. 10
  • The 30-day pair-level logistic model uses pre-t0 user covariates and first-event and scope indicators; confidence intervals cluster by user, not shared authors.Table 3 lists the user covariates, each computed from events preceding the pair’s first exploratory contact.Source: A Robustness, Validation, and Model… · pp. 10–11, Table 3 · p. 11
  • Adding target-author connectivity shifts the distant-contact odds ratio from 0.351 to 0.289, without explaining away the association with subsequent tie formation.The target covariates include author in-degree, out-degree, and received-interaction volume at first contact.Source: A Robustness, Validation, and Model… · p. 11
  • Across 15 overlapping user subsets, distant-contact odds ratios remained between 0.331 and 0.380, with every interval excluding one.The subsets vary by activity, degree, tenure, exploration rate, and position in the observation window.Source: A Robustness, Validation, and Model… · p. 11
  • Replacing scope with post-contact event-sequence measures yielded a local-event-share association of 1.682, below the 2.849 redundancy benchmark.Among pairs contacted again, the local share was 0.834 for local-first pairs and 0.073 for distant-first pairs.Source: A Robustness, Validation, and Model… · p. 11
Loading 2610.01971v1…