fg-zettelkasten

Bridging the narrative divide: Cross-platform discourse networks in fragmented ecosystems

4 min read

Bridging the narrative divide: Cross-platform discourse networks in fragmented ecosystems

Summary

This paper proposes Cluster Affiliation Network Embedding (CANE) and a temporal variant t-CANE, a platform-agnostic method for building user-user networks from shared participation in latent narrative clusters rather than from reposts, follows, or other platform-specific signals. By embedding posts with MPNet, clustering them with DP-Means, and linking users via TF-IDF-weighted affiliation vectors (accelerated with FAISS-HNSW), the authors construct discourse networks that match or exceed interaction- and similarity-based baselines on information operation detection, ideological stance prediction, and a newly introduced cross-platform engagement prediction task — using a fraction of the data. Applied to Truth Social and X during the 2024 U.S. Presidential election, the framework uncovers a small set of “bridge users” (0.33% of users) who introduce roughly 70% of narratives that migrate between the two platforms, reframing cross-platform influence as a structural phenomenon rooted in discursive alignment rather than direct ties.

Key Contributions

  • A content-driven, platform-agnostic framework (CANE / t-CANE) for reconstructing user networks from latent narrative participation, bypassing reliance on increasingly inaccessible interaction data.
  • A novel cross-platform engagement prediction benchmark that tests whether discourse-based graphs encode behavioral alignment across fragmented ecosystems.
  • Empirical evidence that political narrative diffusion between Truth Social and X is structured, directional, and concentrated in a small bridge population.
  • A theoretical link between computational discourse-network analysis and classical sociological constructs — Granovetter’s weak ties, Gould–Fernandez brokers, and boundary spanners.
  • Released anonymized code, datasets, and a cross-platform Truth Social/X corpus.

Methods

Posts are embedded with MPNet and clustered into latent narratives via DP-Means (cosine threshold ≈0.65). Users are represented as TF-IDF-weighted distributions over clusters; user-user edges are computed via cosine similarity with FAISS-HNSW nearest-neighbor search. The temporal extension t-CANE applies a Hawkes-process-inspired decay over discrete timesteps. Evaluation spans (i) state-backed information operation detection on China/Iran datasets, (ii) ideological stance prediction on labeled X and TikTok 2024 election data using GCN/node2vec classifiers, and (iii) a new cross-platform engagement task over 321 narrative themes from May–Nov 2024 Truth Social/X posts. Cross-platform diffusion is analyzed with Transfer Entropy (permutation-tested), Louvain communities, Shannon entropy for identifying mixed-platform bridge zones, and a fear-speech classifier for content analysis.

Findings

  • t-CANE achieves state-of-the-art results across all tasks (e.g., F1 0.83/AUC 0.98 on China IO; F1 0.35/AUC 0.94 on cross-platform engagement vs. best baseline F1 0.11/AUC 0.64).
  • CANE/t-CANE reach 95% of peak AUC using only 5–10% of available content, dramatically outperforming baselines in data efficiency.
  • Of 1,552 cross-platform migrating narratives, 238 (15.3%) show statistically significant directional diffusion via Transfer Entropy; Truth Social is overrepresented as origin by ~11–14× relative to its post share.
  • Truth-Social-originating narratives contain ~22.5% more fear-laden language than X-originating ones (log-odds +0.22, p<0.01).
  • A single high-entropy community of 2,864 users (0.33% of users, 2.14% of posts) is the first carrier of ~68–69% of migrating narratives; 122 users account for all earliest introductions of Truth Social narratives into X, with 4 users responsible for ~25%.
  • Narratives where bridge users engage early (first 5% of participants) receive significantly higher likes, replies, and reposts than matched controls — robust even for low-virality narratives.
  • Ablations confirm TF-IDF affiliation weighting beats raw counts/softmax, and FAISS-HNSW matches brute-force similarity at far lower cost.

Connections

This work sits squarely within current efforts to study coordinated and cross-platform information dynamics under deteriorating API access, alongside Minici2024-tf and Luceri2025-tr on coordination detection, and Lai2024-to and Freelon2024-sc on inferring structure from limited signals. Its cross-platform narrative-diffusion framing connects to Bouchaud2026-lr, Bastos2025-ya, and Bastos2025-ol on multi-platform political communication, while the methodological turn toward latent semantic/topic representations of users echoes Balluff2026-if and Kansaon2025-id. The “bridge user” finding resonates with broader work on small influential subpopulations such as Bak-Coleman2026-mk and Graham2025-gp on structurally pivotal accounts in online discourse.

Podcast

A research-radio episode discusses this paper: Listen

Graph View

Backlinks