CollabNET: Interactive Bibliometric Collaboration & Social Network Analysis

CollabNET is a high-performance, client-side bibliometric and social network analysis platform compiled to WebAssembly (WASM via Shinylive). It empowers researchers, bibliometricians, and research institutions to analyze co-authorship networks, research impact, interdisciplinary intersections, and topological graph-theoretic properties directly inside the web browser—without requiring a dedicated analysis server for browser-side file workflows. Live metadata lookups and optional AI integrations require network access; privacy depends on which external services and AI providers the user chooses to invoke.

Live application: https://vizkidd.github.io/collabnet/

---

🌟 Key Features

1. Zero-Backend Ingestion & Portability

• Live Resolvers: Instant asynchronous publication and metadata extraction via ORCiD (v3.0 REST API), Scopus Author ID / Document Search, OpenAlex, and Crossref.
• Universal File Ingestion: Intelligent column mapping wizard for Scopus CSV, Web of Science (WoS), PubMed, Lens.org, and custom CSV/Excel files.
• Session Persistence (.cnet v2): gzipped JSON with a schema version. It holds the data table, source metadata, retrieval time, filter, counting, community and layout settings, and the app/R/igraph versions. It is validated strictly on load. Legacy binary (RDS) .cnet/.rds files are refused, because R deserialisation of untrusted files can execute code.
• Tabular Exports: Publication tables exportable to CSV, Excel, PDF, and clipboard via DataTables Buttons.

2. Rigorous Graph Theory & Network Topologies

• Newman Fractional Weighting: Edge weights between co-authors scaled inversely by the number of authors per publication: $$w_{ij} = \sum_{p \in P_{ij}} \frac{1}{N_p - 1}$$
• Centrality Suite: Node-level Degree, Betweenness, Closeness, Eigenvector centrality, PageRank, and k-Core decomposition.
• Hamiltonian Graph Diagnostics: Automated validation of Ore's Theorem and Dirac's Theorem for Hamiltonian cycle existence in collaboration components:
  • Ore's Condition: $\deg(u) + \deg(v) \ge |V|$ for every pair of non-adjacent vertices.
  • Dirac's Condition: $\deg(v) \ge \frac{|V|}{2}$ for all vertices.
• Cut-Vertex & Articulation Point Detection: Identifies critical collaboration gatekeepers whose departure would partition the network.
• Dynamic Ego Networks: One-click egocentric neighborhood extraction for selected authors.

3. Boolean Relationship Gates

Combine multiple focal authors or lookup keywords using propositional logic gates:

• OR (Co-patriot/Collaborator): Publications containing any selected keyword.
• AND (Companion): Publications where all focal entities co-authored together.
• XOR (Rival): Publications authored by strictly one focal author (zero overlap).
• NOR (Ignore): Publications authored by collaborators excluding the focal authors.
• NAND (Divide & Exclude): Exclude publications that feature all focal entities simultaneously.
• FULL (Unfiltered): Full dataset preserved with target authors annotated.

4. Overlap & Interdisciplinary Analysis

• High-resolution Venn diagrams and UpSet-style intersection plots across keywords, research departments, or publication venues.
• Dynamic publication and citation contribution trends stratified by OpenAlex citedness bands (Q1–Q4 by fixed cut-offs on 2-year mean citedness: ≥4, ≥2, ≥0.75, >0; not JCR quartiles) and authorship position (First, Second, Corresponding, Co-Author).

5. Single-threaded by design

• All computation runs in webR (single-threaded R in WebAssembly). An earlier JavaScript Web Worker Pool was never wired to R and was removed in 2.2.0 (P2-6); large graphs rely on exact-but-efficient igraph routines and opt-in approximations.

🏗️ Architecture

┌──────────────────────── Browser tab ────────────────────────┐
│ Shinylive / webR (single-threaded R in WebAssembly)         │
│  • Shiny reactive engine      • igraph graph engine          │
│  • visNetwork rendering       • DataTables UI tables         │
│  • MCP registry (GScholarLENS-MCP.R) ⇄ collabnet-webmcp.js   │
└──────────────────────────────────────────────────────────────┘

CollabNET operates as a standalone Single Page Application (SPA). Shinylive loads an in-browser webR engine compiled with Emscripten. Imported data and graph calculations are handled in the browser session. However, optional metadata resolvers contact their respective external APIs, and optional AI features send the context/results selected by the user to the configured provider. Do not assume every workflow is offline or that no information leaves the browser.

---

🚀 Quick Start

Online (Recommended)

Navigate to https://vizkidd.github.io/collabnet/ in any modern desktop browser (Chrome, Firefox, Safari, Edge).

Local Browser Hosting

1. Clone the repository:

   git clone https://github.com/vizkidd/collabnet.git
   cd collabnet
   Rscript -e "source('buildWASM.R')"

2. Serve the wasm_build folder with any local web server:

   python3 -m http.server 8000 --directory wasm_build

3. Open http://localhost:8000 in your browser.

---

🤖 Model Context Protocol (MCP) Integration

CollabNET's MCP server runs inside the open browser tab and works on your live session: there is no sample or static data mode. One tool registry (GScholarLENS-MCP.R, 22 tools, plus collabnet_r_eval in debug mode) serves:

• the built-in AI chat (in-page API window.collabnetMCP);
• in-browser agents through WebMCP (document.modelContext / navigator.modelContext, feature-detected; a draft API);
• desktop agents (Claude Desktop, Cursor, VS Code) through one zero-dependency local bridge, mcp/collabnet-mcp.mjs (Node ≥ 18; stdio + Streamable HTTP on 127.0.0.1).

Quick start: node mcp/collabnet-mcp.mjs --print-config → paste into your client → open the one-click link the bridge prints → switch on AI access in Connect AI.

AI access is off by default and enforced in R. Mutating tools follow a consent policy (ask every time / allow read-only automatically / allow all for 30 min), and there is a kill switch. The optional AI chat and any connected agent receive the data their tools return; the chat sends it on to the model endpoint you configure. See README_MCP.md for setup, the tool list, the back-compat name map, the threat model and troubleshooting, and MCP-AUDIT.md for the audit that motivated this design.

Debug mode (top bar) opens a floating R console running in the app's webR session. AI access to R (collabnet_r_eval) exists only while Debug and the console's "Allow AI access" switch are both on.

📚 Documentation & Manual

For full mathematical formulations, tutorial walkthroughs, and step-by-step guidance, refer to the MANUAL.md or click the Manual button in the dashboard top navigation.

📄 License

This project is licensed under the MIT License. See LICENSE for details.

Sh-index author-position weights (C1-4)

The Sh-index is the h-index of position-weighted citations (each paper's citations × the weight of the target author's position on it). Defaults (GScholarLENS): first author 0.90, second 0.50, corresponding/last 1.00, middle co-author 0.25 on papers with ≤ 6 authors or 0.10 with > 6, keyword match without known position 1.00, not found 0. In Collaboration Metrics → Sh-index author-position weights you can view and edit them (0–2, Reset to defaults). The toggle Newman weights for all positions: 1/(n−1) gives every position 1/(n−1), n = number of authors on the paper; single-author papers (n = 1, where 1/(n−1) is undefined) get weight 1. Note that Newman's 1/(n−1) is a pairwise-tie weight (fractional per-author credit would be 1/n); it is applied per position here as requested. Whenever the weights are not the defaults (or Newman is on) a warning banner says which weights differ and that the results are not comparable to standard Sh-index values. The weights are saved in .cnet sessions and stated in the methods paragraph.

Loading a .cnet session: data only or full replay (C1-5)

When an upload contains a .cnet file, the import dialog asks how to load it. Load data only (default) imports the table and provenance and rebuilds with your current settings. Replay full session also pushes every saved setting back into the controls (column, identity policy, edge weights, author-count window, filters, year range, community algorithm/seed/resolution, betweenness mode, layout options, Sh-index weights and Newman toggle): controls are frozen while they are updated so the network is rebuilt once. A notification lists what was applied and any saved setting that could not be applied (no UI control, or a session saved by another CollabNET version). Legacy R-serialised (v1) sessions are still refused (6A security decision).

Native browser smoke test (C1-3)

tests/browser/smoke_native.py runs a minimal native Shiny harness (tests/browser/harness_app/, real www/js + console + Sh-weights panel) in headless Chromium via Playwright and fails on any page error; it does not exercise webR/Shinylive. Run: RSCRIPT=Rscript python tests/browser/smoke_native.py --shots /tmp/shots (add --js-dir <old js> --expect-crash to reproduce the pre-fix crash).

Full-app browser smoke and layout benchmark (C2-0 / C2-1)

tests/browser/smoke_full.py runs the full app natively (shiny::runApp) in headless Chromium and walks upload → network → console → Sh banner → worker layout → .cnet save/reload; it records console/page errors, CSP violations and peak RSS. Status: page load verified with zero errors; the scripted upload step is not yet reliable (see FIXES_C2.md). tests/browser/bench/run_bench.py <vis-network.min.js> 500,2000,5000 vis_physics,server_coords,worker,main_kernel out.json measures long tasks / total blocking time per layout mode. Neither tests webR/Shinylive.

Export wait and optional support (E1/E3)

When data is loaded, each export waits 5 minutes and then downloads for free. You can optionally support the developer with a voluntary gift of any amount on Buy Me a Coffee to skip the wait for the rest of the browser session. On a Cloudflare deploy the dialog shows a one-time code (e.g. CN-7KQ4X2) to put in the BMC message; the tab unlocks when BMC's signed webhook arrives ("Check by email" and "Already a monthly supporter? Check" are fallbacks). Where no verification Worker is reachable (GitHub Pages, local runs) an "I've donated" honour button is shown instead (options(collabnet.gate_honour_fallback = "auto"|"on"|"off")). A gift is not a purchase of data, a service or a licence; no 80G or other tax receipt; refunds only under BMC's refund policy (a refund re-enables the wait). See deploy_static/donation-policy/, GATE_INVENTORY.md, DEPLOY_CLOUDFLARE.md and BMC_WEBHOOK_NOTES.md. Build/deploy targets: Rscript buildWASM.R --target=ghpages|workers|pages|both.