❗ Please note that this is a vibe-coding project accompanied by a vibe-paper here: https://P0L3.github.io/reference_graphs/ . Every statement is to be taken with a grain of salt 
; still, this is good inspiration for what can be done with multiplex networks. ❗
A pipeline for building and visualising multiplex organisational networks from Croatian academic data. It pulls publication and project records from the public CROSBI and CroRIS APIs, resolves researcher identities across sources, and assembles a four-layer graph that can be explored in 3D, analysed statistically, or exported to Gephi/Cytoscape.
Built around the faculties of the University of Rijeka — FIDIT, FABRI, FZF, and FM — but easily extended to any Croatian institution with a CROSBI ID via a single config file (see Configuration below).
Every institution/faculty-related constant used across the pipeline — the roster, the derived dataset list, colors, markers, and the "unaffiliated researcher" sentinel label — lives in one place: config.py. No other script defines its own copy of any of these anymore; they all import from here. Add, remove, or re-color an institution once, and it flows through fetching, building, every RQ module, and every plot without touching those files.
# config.py
INSTITUTIONS = [
{"name": "FIDIT", "crosbi_id": 289, "mbu": 318},
# add or replace entries here — crosbi_id/mbu are required for the CROSBI/CroRIS APIs
...
]
INSTITUTION_COLORS = {
"FIDIT": OKABE_ITO["sky_blue"],
...
EXTERNAL_LABEL: "grey",
LEGACY_UNKNOWN_LABEL: "lightgrey",
}
INSTITUTION_MARKERS = {
"FIDIT": "^", # Triangle Up
...
EXTERNAL_LABEL: "o", # Circle (fallback)
LEGACY_UNKNOWN_LABEL: "o",
}DATASETS and COMBINED_DATASET_NAME (e.g. "FIDIT_FABRI_FZF_FM") are derived automatically from INSTITUTIONS rather than typed out separately, so they can never drift out of sync with the roster above.
fetch_crosbi_data.py → nodes / edges / keywords CSVs
↓
build_graph_crosbi_data.py → GraphML + GEXF per layer (for Gephi/Cytoscape/PyVis)
↓
├── [Primary] run_pipeline.py → Aggregated MASTER comparison CSVs in RESULTS/
│ (builds every dataset in-memory via builder.py, then calls the
│ rq1-rq4 functions below internally — this is the script you run)
│
└── [Optional / diagnostic]
multiplex_graph_analysis_crosbi_data.py → interactive 3D plot & ego-network CLI
All of the above pull their institution list, dataset list, colors, and markers from config.py (see Configuration).
If you only run one script, run this one.
python run_pipeline.pyIt automatically builds every dataset defined in config.py in memory (each institution alone — by default FIDIT, FABRI, FZF, FM — plus the combined university, FIDIT_FABRI_FZF_FM), runs the full RQ1–RQ4 analytical suite on each, and writes tagged, comparable master tables to RESULTS/ — one row per Dataset per metric, so a single institution can be compared directly against the others or against the whole university. This is the setup used to produce every comparative figure in the paper.
rq1_script.py through rq4_script.py (below) are not separate pipeline steps — they are the individual analysis modules that run_pipeline.py imports and calls internally (rq1_process_graph(), rq2_process_graph(), etc.). You do not need to run them yourself. They're documented separately only so that:
- you can inspect exactly which metric belongs to which research question, and
- you can run a single one standalone against one dataset while debugging or drafting a specific figure, without re-running the whole batch.
| # | Layer | Type | Edge = |
|---|---|---|---|
| 1 | Co-authorship | Undirected | Two researchers share a publication |
| 2 | Mentorship | Directed (mentor → student) | Supervisor–student relationship on a thesis (skipped in Co-authorship if student thesis) |
| 3 | Research Similarity | Undirected | Pairwise Jaccard similarity |
| 4 | Project Co-participation | Undirected | Two researchers appear on the same funded project |
REF_GRA/
├── config.py # ⚙️ Single source of truth: institutions, dataset list, colors, markers
├── fetch_crosbi_data.py # Step 1 — API extraction & entity resolution
├── build_graph_crosbi_data.py # Step 2 — Graph assembly, export for external tools
├── builder.py # In-memory multiplex graph constructor, used by run_pipeline.py
├── run_pipeline.py # ⭐ MAIN ENTRY POINT — master automation engine, all datasets
├── rq1_script.py # Module: Baseline topologies & Null models (called by run_pipeline.py)
├── rq2_script.py # Module: Multiplexity, triads & edge robustness (called by run_pipeline.py)
├── rq3_script.py # Module: Aggregation cost & interdisciplinary brokers (called by run_pipeline.py)
├── rq4_script.py # Module: Core-periphery & Community persistence (called by run_pipeline.py)
├── collab_recommender.py # Actionable insights: Synergetic recommendations
├── multiplex_graph_analysis_crosbi_data.py # Optional — Interactive 3D layout & ego-network CLI
├── data_stats.py # Calculate publication and project statistics.
├── plot_master_results(_2)/visuals.py # Plot the Figures used in the related paper (read from RESULTS/)
├── crosbi_coauthorship_institution.ipynb # Interactive notebook (older, single-institution)
├── DATA/ # Raw pickles / CSVs (not tracked)
├── RESULTS/ # Aggregated multi-institutional CSV analysis tables (from run_pipeline.py)
└── exported_graphs/ # GraphML & GEXF outputs (generated at runtime)
conda create -n refgra python=3.10
conda activate refgra
conda install pandas networkx matplotlib pyvis ipykernel requests tqdm nbconvert scipy scikit-learn seaborn pymnet -y
# python -m pip install pymnetpython fetch_crosbi_data.pyQueries the CROSBI publications API and the CroRIS projects API for every institution defined in config.py, resolves researcher identities using a three-tier strategy (project metadata
nodes_FIDIT_FABRI_FZF_FM.csv # one row per unique researcher with institution
edges_FIDIT_FABRI_FZF_FM.csv # raw co-authorship, mentorship, project edges
keywords_FIDIT_FABRI_FZF_FM.csv # keyword associations per researcher per publication
To analyse a different set of institutions, edit INSTITUTIONS in config.py — fetch_crosbi_data.py imports this list rather than defining its own copy, so this is the only place you need to make the change:
# config.py
INSTITUTIONS = [
{"name": "FIDIT", "crosbi_id": 289, "mbu": 318},
# add or replace entries here
]python build_graph_crosbi_data.pyLoads the three CSVs, computes research-similarity edges via pairwise Jaccard comparison on keywords, assembles four NetworkX graphs, and exports each layer as both GraphML and GEXF under exported_graphs/. This step is for feeding Gephi/Cytoscape/PyVis — it is not required by run_pipeline.py, which builds its own graphs in memory via builder.py.
python run_pipeline.pyThis is the script you should run for any statistical/comparative analysis. It iterates through every dataset in config.py — each institution in isolation (FIDIT alone, FABRI alone, FZF alone, FM alone) as well as the combined global university graph — constructs each graph in-memory via builder.py, runs all four RQ modules on each, and writes tagged master comparison tables to RESULTS/:
RESULTS/MASTER_RQ1_Topologies.csv
RESULTS/MASTER_RQ2_EdgeMultiplicity.csv
RESULTS/MASTER_RQ2_TriadicClosure.csv
RESULTS/MASTER_RQ2_LayerSimilarity.csv
RESULTS/MASTER_RQ2_StructuralTwins.csv
RESULTS/MASTER_RQ3_AggregationStats.csv
RESULTS/MASTER_RQ3_NodeMetrics_Full.csv
RESULTS/MASTER_RQ3_NodeMetrics_Ranked.csv
RESULTS/MASTER_RQ4_CommunityPersistence.csv
RESULTS/MASTER_RQ4_CoreDecomposition.csv
RESULTS/MASTER_RQ4_CoreStats.csv
RESULTS/MASTER_RQ4_EliteCoreBreakdown.csv
RESULTS/RUN_MANIFEST.csv
Every table has a Dataset column so any metric can be compared across FIDIT / FABRI / FZF / FM / the combined university. RUN_MANIFEST.csv records which datasets were successfully processed vs. skipped (e.g. missing CSVs, or a dataset that errored partway through) so a partial run is always auditable rather than silently incomplete.
Ideal for testing macro (whole-university) vs. micro (single-faculty) organisational cultures.
python multiplex_graph_analysis_crosbi_data.pyLaunches an interactive CLI to explore one dataset's network visually — useful for spot-checking or preparing a qualitative figure, not part of the statistical pipeline above.
Prompts:
1. Researcher name substring search (accent-insensitive, e.g. 'mestrovic')
2. Number of high-centrality context nodes to render alongside (default 120)
The output is an egocentric multiplex view: four horizontal planes (one per layer), cross-layer vertical coupling lines, and node positions anchored across planes.
Visual encoding
| Element | Encoding |
|---|---|
| Plane colour | Layer identity |
| Node shape | Faculty affiliation, per INSTITUTION_MARKERS in config.py (▲ FIDIT · ★ FABRI · ■ FZF · ◆ FM · ● External) |
| Node position | Unified 2D coordinates projected onto vertical Z-planes |
| Node opacity | Primary interaction layer = 0.90 · Ghost layers = 0.07 · Background = 0.15 |
| Edge weight | Thicker / black for edges incident to the target researcher |
| Vertical dashed line | Cross-layer coupling; bold for the ego node |
You don't need to run these directly; they're documented here so each metric can be traced to the research question it answers. Each also has a standalone __main__ block for debugging a single dataset in isolation.
Calculates density, component count, and GCC. Tests institutional homophily (Assortativity) against Degree-Preserving Null Models to test whether departmental silos are driven by sociology or random chance.
Computes Edge Multiplicity (robustness of ties) and the Layer Similarity Matrix. Tests Multiplex Triadic Closure to see if social capital in one layer (e.g., Projects) successfully transfers into another (e.g., Publications).
Calculates the Spearman Rank Correlation between Monoplex and Multiplex PageRank to quantify the percentage of information destroyed when flattening a network. Identifies interdisciplinary boundary-spanners using the Participation Coefficient (
Uses Multilayer
python collab_recommender.pyScans the network for researcher pairs who share a high thematic keyword similarity but have zero co-authored publications or shared projects. Generates a ranked checklist of inter-departmental synergy opportunities.
Metric/measure implementation details reference guide is available here.
| Source | API base | What it provides |
|---|---|---|
| CROSBI | https://www.bib.irb.hr/crosbi-api/ |
Publications, authors, keywords, mentorship |
| CroRIS | https://www.croris.hr/projekti-api/ |
Funded projects & team members |
Both APIs are public and unauthenticated. Rate-limit your requests if running against a large institution.
- Migration to pymnet with a implementation guide.