PaperlayerView exact revision on arXiv

Source-linked AI summary

Who Belongs Together? Topical and Social Structure in Bluesky Starter Packs

Sima Adleyba, Onur Varol

arXiv:2609.31297v1cs.SIcs.CY

TL;DR

Whether starter packs capture meaningful topical and social structure remains an open question, despite their use in helping newcomers discover accounts. The study examines packs using metadata, posts, inferred demographics, and follow networks. It finds coherent topical and social groupings, while noting limits in distinguishing selection effects from pre-existing differences.

  • Problem

    Whether starter packs capture meaningful topical and social structure, and how their membership relates to curators and existing follow relationships, remains unanswered.

    Source: Introduction · pp. 1–2

  • Method

    The study analyzes 2,718 active starter packs by combining pack metadata, account post histories, follow networks, and inferred demographic attributes.

    Source: Introduction · p. 2

  • Results

    Starter packs capture semantically coherent subgroups and social structure, with stronger member–curator content resemblance than comparable same-topic accounts and demographic resemblance varying by attribute and topic.

    Source: Conclusion · p. 8

  • Takeaways & Limitations

    Because a typical member’s audience follows only a small fraction of the pack, relevant pack recommendations could expose users to accounts beyond their immediate follow networks.

    Source: Conclusion · p. 8

    The study cannot distinguish pre-existing differences from changes associated with pack inclusion, and future work should test whether findings extend to less-adopted packs.

    Source: Conclusion · p. 8

Abstract

from arXiv · show

Bluesky starter packs are human curated collections of accounts and feeds aimed at helping users discover new communities, especially during onboarding. Prior research has examined their effects on platform growth and account visibility. However, whether these packs capture meaningful topical and social structure remains unanswered. We study 2718 active starter packs containing approximately 144K accounts, combining pack metadata, follow graphs, user post histories and inferred demographic attributes to answer this question. We assign starter packs to 17 topical categories, and examine semantic coherence, curator--member resemblance, shared audiences, and follow reciprocity. We found that pack members are closely aligned with both their fellow pack members and their pack's description compared to accounts included in other packs of the same topic, indicating topical specificity beyond the broader topical categories. We also identify semantic resemblance between curators and their pack members, while demographic resemblance varies by attribute and topic. Network analyses show higher follow reciprocity among members of the same pack and between members and curators than between a pack member and either another member from the same broader topic or a nonmember. Members also share substantial portions of their follower networks with fellow pack members. Despite this coherence, a typical member's follower follows only a small fraction of other members in the pack, indicating limited audience coverage. These results suggest that starter packs capture coherent topical and social groups while leaving opportunities for further account discovery. More broadly, results show human-curation can be useful to capture relevant communities and connections, and platforms could integrate human-curation into recommendation systems as a complement to algorithmic systems to expand opportunities for discovery.

Introduction

Platform migration can leave users rebuilding connections, while starter-pack curation shapes which accounts newcomers see. This study asks whether packs capture topical and social structure and finds alignment within packs beyond broad topic labels.

Introduction

  • Starter packs make deliberate choices about who is grouped together and exposed to newcomers, raising questions about topical fit, curator resemblance, and follow-network ties.The study examines these questions across 2,718 active packs and 144,987 unique accounts.Source: Introduction · pp. 1–2
  • The study combines pack descriptions, account posting histories, follower networks, and inferred demographic attributes to examine three questions about alignment, curators, and member connections.Source: Introduction · p. 2
  • Members align more closely with their own packs than same-topic accounts, and pack members are more semantically similar than random accounts from the same broader topic.Demographic resemblance varies across attributes and topics, while members share audiences and show higher follow reciprocity with fellow members and curators.Source: • RQ2: Do pack curators resemble the… · p. 2
  • Starter packs capture topical and social structure within broad topics, suggesting potential for discovery beyond onboarding and curator-shared links.The paper proposes that Bluesky could suggest relevant packs to users who already follow one or more of their accounts.Source: • RQ2: Do pack curators resemble the… · p. 2

Related Work

Prior work links platform migration to social connections and describes starter packs as an onboarding aid, while research on curated lists motivates examining their topical, social, and demographic structure. Work on unequal visibility further frames pack selection as relevant to exposure.

Related Work

  • Migration research emphasizes the role of existing social connections, while studies of Bluesky’s growth describe an opportunity to examine how newcomers discover accounts and build networks.Source: Platform migration and the growth of… · p. 2

Cold start and starter packs

  • Starter packs offer a non-algorithmic way to recommend accounts and feeds together, extending onboarding work on imported connections that do not replace organic discovery.Prior research reported that starter packs accounted for up to 43% of daily follow operations at their peak and linked inclusion to follower growth and activity.Source: Cold start and starter packs · p. 2

Human curation and social groups

  • Research on human-curated lists shows they can reveal topical expertise and communities, while curated-group studies connect membership to social structure and ties beyond the group.Work on homophily motivates measuring curator–member semantic and demographic resemblance separately.Source: Human curation and social groups · pp. 2–3

Bias and fairness

  • Because visibility varies across demographic groups and popularity, pack composition matters; only 18 of 2,718 analyzed packs exceed 150 members, versus 265 of 273,524 in the full collection.Prior work also links recommendation algorithms to demographic underrepresentation and notes that pack inclusion can complicate comparisons between members and nonmembers.Source: Figure 1 · p. 3, Bias and fairness · p. 3

Data and Methods

The study filters a large collection of starter packs, assigns packs to topical categories, and constructs account and pack representations from posts and metadata. It then checks semantic similarity patterns and temporal stability, noting limits in coverage and interpretation.

Data collection

  • From 273,524 fetched packs, the analysis retains 2,718 with at least five followers, covering 144,987 unique accounts and 192,491 pack-member edges.Follow records were returned for 142,972 accounts, yielding approximately 103 million directed follow edges.Source: Data collection · pp. 3–4

Assignment of pack communities

  • The authors assign packs to one of 17 fixed topic labels using Gemini 3.5 Flash after BERTopic produced generic clusters; one author’s 200-pack validation sample yielded Cohen’s κ = 0.912.The pack-overlap network connects packs whose member overlap coefficient meets the 0.05 threshold.Source: Assignment of pack communities · p. 4

Embeddings for members and packs

  • Embeddings represent packs with title-and-description text and accounts with post vectors, averaging eligible post vectors to form accounts and member centroids.Originals, reposts, and quotes are embedded, but replies are excluded; media is omitted and 88.8% of posts produce a vector.Source: Embeddings for members and packs · pp. 4–5
  • Mean cosine similarity rises from 0.789 for random pairs to 0.832 within a topic and 0.909 within a pack, supporting topic differentiation and stronger within-pack similarity.Among same-topic pairs, those sharing a pack average 0.907 versus 0.831 for those that do not; the ordering holds within every topic and with original posts alone.Source: Embeddings for members and packs · p. 5
  • Temporal agreement is high among the 75,194 accounts meeting the post threshold, but the authors caution that it does not establish unchanged behavior or cover excluded accounts.Mean and median similarity between first- and last-window centroids are 0.979 and 0.985; agreement also rises with post count, so volume matters when interpreting cosine results.Source: Embeddings for members and packs · p. 5

Results

Starter-pack members align with their own pack more closely than same-topic comparison accounts, and they also resemble their curators semantically. Members share follower networks and reciprocate follows more within packs, but pack fullness remains low, leaving much of each pack outside many followers’ networks.

Results

  • Member alignment AUCs range from 0.754 to 0.954 and description alignment AUCs from 0.725 to 0.898 across 17 topics, exceeding same-topic comparison accounts.Together with pairwise similarity comparisons, this indicates topical specificity beyond broader topic labels.Source: RQ1: Are starter-pack members more… · p. 6
  • Across all 17 topics, members resemble curators more semantically than same-topic comparison accounts, with mean AUCs from 0.725 to 0.937.Gender skews are positive in every topic, while age and race vary in magnitude, direction, and significance; the analyses do not establish whether resemblance preceded or followed pack inclusion.Source: RQ2: Do pack curators resemble the… · p. 6
  • Average follower overlap ranges from approximately 0.35 to 0.60 across topics, indicating that pack members share a considerable portion of their follower networks.NSFW / Adult Content, Artists, and LGBTQ+ Community show higher content similarity alongside lower overlap, while Educators (K-12) and Sports show higher overlap.Source: RQ3: How cohesive and complete are the… · p. 7, Figure 5 · p. 8
  • Within-pack reciprocity exceeds both analysis-wide member reciprocity and nonmember reciprocity; in most topics, curator follow rates are higher still.The pattern indicates stronger reciprocal connections among members of the same pack than with nonmembers.Source: RQ3: How cohesive and complete are the… · p. 7, Figure 5 · p. 8
  • At least half of packs in every topic have fullness below 50%, and the corpus-wide median is approximately 0.21.Fullness is calculated among followers who follow at least five pack members; its strong negative correlation with pack size (ρ = −0.849) should be interpreted in light of the lower bound 5/|Mp|. Low fullness means much of the pack remains outside those followers’ networks, leaving potential for discovery.Source: RQ3: How cohesive and complete are the… · pp. 7–8, Figure 6 · p. 8

Conclusion

Starter packs capture semantically coherent subgroups within broader topics, with curator resemblance and shared audiences also indicating topical and social structure. Members’ audiences typically follow only a small fraction of the other pack members, leaving room for further discovery.

  • Within broader topic areas, starter packs capture semantically coherent subgroups, and members resemble their curators in posting content; demographic resemblance is positive for some attributes and topics.Source: Conclusion · p. 8
  • Shared followers and high follow reciprocity among members and between members and curators indicate that packs capture social as well as topical structure.Source: Conclusion · p. 8
  • A member’s audience typically follows only a small fraction of the remaining pack members, while future work could test discovery recommendations and whether findings extend to less-adopted packs.Source: Conclusion · p. 8

Prompt A.1: Starter pack topic detection template

The template directs a classifier to assign starter packs to relevant topics using the pack name and description. It specifies topic choices, confidence fields, and a JSON response format.

Loading 2609.31297v1…