About This Knowledge Graph

Alymani, A. A., Alsofiani, M., Murugesan, L., & Alammar, A. (2026). GML/GNN in AEC: A PRISMA 2020 Knowledge Graph and Wiki (Version 1.0.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.20073214

This site presents a knowledge graph and wiki for 112 peer-reviewed papers on Graph Machine Learning (GML) and Graph Neural Networks (GNNs) applied across Architecture, Engineering and Construction (AEC), identified through a PRISMA 2020 systematic literature review (search dates: February–March 2026).

Data Sources

• 35 CumInCAD seed papers — primary AEC architectural-computing database; full-text screened.
• 35 backtracking papers — citation-network expansion from the seeds.
• 42 EBSCO papers — Applied Science & Technology Source Ultimate, Academic Search Ultimate, Springer Nature Journals, IEEE Xplore, ScienceDirect, and Emerald Insight, accessed via SDL institutional access.

Five Metadata Layers

Every paper is described in a 5-layer JSON schema. Each layer is independently consumable; the wiki page renderer merges all five into a single view.

1. Bibliographic — Authors, year, title, venue, DOI (with source provenance), link, source database.
2. Categorical — Bridge-to-GNN category (A–E), Excel category, PDF location.
3. Methodological — GNN architecture, graph encoding type, primary AEC task, learning task, graph type. Inferred from title + category and tagged with a (detected from PDF) label when surfaced from the deep extraction.
4. Relational — Membership in 8 named lineage families, citations, extended-by relations, infrastructure dependencies, co-author overlaps. Auto-suggested memberships are marked with a yellow (auto-suggested) tag.
5. Temporal & Quality — Cohort, frontier tags, gap indicators, rigor score (6-criterion checklist), reproducibility flags.

9 Connection Modes (typed edges in the graph)

The interactive graph at /graph.html renders any one of these modes at a time, then redraws with anchored physics so the layout stays comparable across modes.

1. Citation / Lineage — directed, hand-curated; 8 named families (Floorplan Generation, Topologic Infrastructure, Energy Digital Twin, BIM Classification, Structural-Frame GNN, Safety & KM, LLM-KG Hybrid, BIM-GIS Integration).
2. Category Cluster — same Excel category; capped at 5 nearest-by-year neighbours.
3. Domain Cluster — same application domain; cap 5.
4. Methodology Similarity — same inferred GNN architecture; cap 4.
5. Infrastructure Dependency — directed, hand-curated; Topologic → users, BOT → followers, Collins-IFC → BIM classifiers, foundational GNN papers → applied work.
6. Temporal Proximity — same publication year; cap 3 cross-category.
7. Community Cluster (NetworkX) — greedy modularity / Louvain communities computed across the entire corpus.
8. Same GNN Architecture (deep-PDF detection) — papers sharing a regex-detected architecture (DGCNN, GraphSAGE, GAT, GIN, R-GCN, MPNN, …) across all 112 full-text PDFs.
9. Shared Framework / Library (deep-PDF detection) — papers sharing detected tooling (PyG, DGL, Topologic, NetworkX, Revit/Dynamo, RDF/SPARQL, …).

5 Layouts (same nodes, same edges, different geometry)

The layout selector only changes node positions; the active connection mode determines which edges are drawn. Switching layouts can make the same edge set look denser or sparser — this is a property of the algorithm, not a change in the data.

1. Spring (force-directed) — D3 force simulation with global repulsion (forceManyBody) and per-edge attraction; spreads nodes across the canvas so edges stretch through whitespace and the graph looks airy, with clusters and outliers physically separated.
2. Kamada-Kawai — deterministic stress-minimisation that packs nodes into a compact, balanced blob with short, near-uniform edge lengths, so the same edge set as Spring appears visually denser (more crossings per square pixel) — same edges, just shorter and overlapping.
3. Hierarchical — BFS layering from the foundational GNN papers downward, useful for tracing influence chains from theory to applied AEC work.
4. Radial — centre-out arrangement keyed to lineage / domain sectors, emphasising group membership over ordering.
5. Temporal — x-axis locked to publication year (anchor strength 0.95) while y-axis drifts (0.30), so the canvas reads left-to-right as a chronology while clusters still emerge vertically within each year.

Provenance System (4 tiers, colour-coded throughout)

• Curated — clean, no tag; manually entered by the review team.
• Manual extraction — teal (extracted) tag; pulled from the methodology Excel template, with per-field confidence.
• Heuristic — yellow (auto-suggested) / (detected from PDF) tag; algorithmic catalogue match across 23 architectures, 33 frameworks, 35 benchmarks, 10 loss functions, 9 activations.
• Bibliometric verification — blue panel; cross-referenced against Connected Papers .bib exports (15 seeds, 20 verified edges).

Build Pipeline

The site is regenerated by a Python pipeline (build/01 through build/05). The single source of truth for the graph layer is data/corpus_graph.json — a master NetworkX MultiDiGraph that drives both the 6 static PNG images on the home page and the D3.js interactive graph. Stats on this site are computed live from that JSON, so they cannot drift.

Acknowledgements

Authored by Abdulrahman Ahmed Alymani, Mohammed Alsofiani, Libish Murugesan, and Ammar Alammar at the CM-iTAD Lab, Alfaisal University, Riyadh, Saudi Arabia. Built with Python, NetworkX, lxml, matplotlib, D3.js, and Vercel.

For the full reproducibility detail, see HANDOFF.md in the GitHub repository.