Skip to content

Coordination Network Toolkit

FreeInstitutional

Publisher: github.com/QUT-Digital-Observatory/coordination-network-toolkit | Type: Non-detector | Cost: Free

TL;DR

A free, open-source R and Python framework from the QUT Digital Observatory for detecting coordinated behaviour through user similarity networks across social-media datasets. The academic- grade alternative in the institutional CIB ladder, with peer- reviewed methodology behind the Copy Pasta Test pattern.

What it does

The Coordination Network Toolkit takes a social-media dataset and constructs a user-similarity network: nodes are accounts, edges are weighted by how similar two accounts' posting behaviour is across chosen dimensions (text re-use, timing, hashtag overlap, link sharing). The output is a network graph that an analyst can inspect for clusters indicative of coordinated amplification. The toolkit's methodology has been published in academic venues including ICWSM 2025, which gives a fact-check operation a citable foundation when its findings are challenged in a defamation suit or platform appeal.

The toolkit operates at the same architectural layer as CIB Mango Tree but with two practical differences. First, it is an R-and-Python framework rather than an interactive Python terminal; it is designed to be embedded in a research-grade analysis pipeline rather than run conversationally. Second, it produces network artefacts (graph objects, edge lists, similarity matrices) suitable for downstream visualisation in Gephi or Maltego, where Mango Tree produces account-cluster lists more directly. Both sit on the behaviour pillar per Architectural Anchor 1.

When to use it

  • A research lab or fact-check operation already has an R / Python analysis pipeline and wants to plug a peer-reviewed coordination- detection layer into it rather than adopt a separate tool.
  • An institutional partner intends to publish a CIB finding and needs a methodology with a citable academic publication record to withstand legal or platform scrutiny.
  • A regional newsroom is collaborating with a university research partner, and the partner's preferred environment is R for statistical work or Python for ML-pipeline integration.
  • A multi-platform investigation needs a similarity construct that can take signals from text re-use, timing, hashtag co-occurrence, and URL co-sharing in a single combined network rather than as separate per-test outputs.

Limitations

Limitations

  • Requires R and / or Python proficiency, plus network-analysis expertise to interpret similarity graphs correctly.
  • Documented use is academic: peer-reviewed papers, not newsroom casework. The toolkit is well-reviewed inside its research community but has no published SEA newsroom deployment.
  • Like every coordination-detection method, similarity networks can surface legitimate political-party social-media operations and authentic communities of interest. Coordination signatures are necessary but not sufficient for an inauthenticity finding.
  • Heavy on input-dataset quality. Sparse data over short time windows produces noisy similarity scores; a meaningful analysis usually wants weeks of post-level data per account.

Privacy and threat model

The toolkit runs on the analyst's machine or institutional infrastructure; no data is sent to a vendor cloud as part of the analysis itself. Privacy considerations therefore concentrate on the input dataset (where it came from, what permissions cover its analysis) and on the output (named-account network graphs, which have the same disclosure-care implications as any other CIB finding).

For collaborations between SEA newsrooms and Western university partners, the data-residency question matters. If the dataset itself sits on a partner's institutional servers, the analysis result inherits that jurisdiction's data-protection regime and any research-ethics constraints attached to the original data. Confirm the chain of custody before publishing a finding generated through a partner's environment.

Country and platform applicability

  • Indonesia: language-agnostic; relevant in principle for Bahasa-Indonesian buzzer-network analysis, especially as a pipeline component in collaborations with academic partners. No published Indonesian deployment.
  • Laos: language-agnostic; the local-execution model is privacy-defensible. No Lao-specific deployment.
  • Malaysia: language-agnostic. No published deployment.
  • Philippines: language-agnostic; relevant in academic partnerships with Filipino universities and the #FactsFirstPH research layer. No published deployment.
  • Sri Lanka: language-agnostic; the LIRNEasia and Hashtag Generation research environment maps onto the toolkit's intended user profile. No documented deployment yet.
  • Thailand: language-agnostic; relevant for institutional partners following up the Anutin / Mauerberger 2026 case at the network-pattern layer.

Platform applicability: works on any platform whose data the analyst can extract. The toolkit's R-and-Python framing makes Meta Content Library extracts the natural input where IFCN access is available.

How to access

Free download from the QUT Digital Observatory's GitHub distribution. Install per the repository README in either an R or Python environment. No account creation, no beta gate. The toolkit is open-source; the maintainer organisation is the QUT Digital Observatory at Queensland University of Technology, Australia.

Cost (current as of 2026-05)

Free. Open-source. The toolkit's continued development depends on QUT Digital Observatory's research funding and on community contributions; it is not exposed to vendor-side pricing changes. The cost of running an analysis is the analyst's time, the local compute for the network construction, and any underlying data- licence costs (Meta Content Library access is gated to IFCN signatories rather than priced commercially).

Quickstart

  1. Confirm whether your institutional pipeline is R-based or Python-based and install the corresponding distribution from the QUT Digital Observatory repository.
  2. Install dependencies per the README.
  3. Prepare the input dataset (post-level records with account ID, timestamp, text, hashtags, URLs as required by the chosen similarity test).
  4. Decide which similarity dimensions to combine: text re-use, timing co-occurrence, hashtag overlap, URL co-sharing.
  5. Run the toolkit's similarity-network construction step.
  6. Export the resulting graph object (edge list, similarity matrix) for visualisation in Gephi or downstream community detection.
  7. Inspect identified clusters for coordination signatures and cross-check against Mango Tree's Copy Pasta Test on the same dataset for triangulation.
  8. Document parameters (time window, similarity threshold, dimensions used) in any published finding so that the analysis is reproducible.

In the toolkit's workflow

Behaviour-pillar non-detector tool per Architectural Anchor 1. Sits in 1C.1 Institutional CIB detection as an academic-grade alternative to CIB Mango Tree. The two are intentionally complementary rather than competing: Mango Tree's interactive Copy Pasta Test is the faster path to a single coordination signal, the Coordination Network Toolkit is the framework where multiple signals combine into one network construct with peer-reviewed methodology behind it.

Standard combinations:

  • With CIB Mango Tree as triangulation: run both on the same dataset and compare cluster boundaries; agreement strengthens the finding, divergence flags methodology sensitivity to parameter choices.
  • With CooRTweet when the analysis is in R and the data source is Meta Content Library; CooRTweet handles MCL data natively.
  • With Gephi at 2C.2 for publication-grade network graphics from the toolkit's exported edge lists.
  • With Maltego at 2C.2 when an entity-graph view is more useful than a similarity-network view.
  • With ABCDE Framework at 2C.3 to codify the behavioural pattern in published reporting.

This card carries no detector signal class: the toolkit produces a non-detector behavioural-pattern signal class. Per Anchor 2, that signal combines with claim-extraction (2B.2) and source-history (1B.1) to support a strong public claim; coordination signatures alone are not enough.

Sources