research-knowledge-graph

Research Knowledge Graphs

A literature review is a graph problem in disguise: which papers everyone builds on, which clusters talk past each other, which work bridges them, and where the gaps are. This skill builds that graph from a reference library and hands it to whichever graph tool the user prefers.

Build

# Offline: authors, venues and keywords from the .bib itself
python scripts/build_graph.py references.bib --out kg/

# Enriched: disambiguated authors (OpenAlex IDs, ORCID), venues, topics, and
# citation edges between papers in the library
python scripts/build_graph.py references.bib --out kg/ --openalex --mailto [email protected]

CSL-JSON (.json, exported by Zotero or reference-manager-interop) also works as input.

Output	Use it with
graph.json	NetworkX: nx.node_link_graph(json.load(f), edges="links"), or any JS graph lib
graph.graphml	Gephi, yEd, Cytoscape, igraph, nx.read_graphml
nodes.csv, edges.csv	Neo4j (neo4j-admin database import full --nodes=nodes.csv --relationships=edges.csv) or Gephi spreadsheet import
vault/	Obsidian (graph view), and Markdown-graph tools such as graphify, which read [[wikilinks]] as edges

Schema. Node type ∈ {paper, author, venue, topic}. Edge type ∈ {authored_by, published_in, about, cites}. Paper nodes carry doi, year, openalex, cited_by_count when known.

Expand beyond the library (snowballing)

Citation edges only connect papers you already have. To map a field:

1. Seed with 5-20 core papers.
2. Backward: collect each seed's referenced_works from OpenAlex. Forward: query https://api.openalex.org/works?filter=cites:W123... for citing works.
3. Keep candidates cited by ≥2 seeds (co-citation), export them to .bib, rebuild.
4. Stop when a round adds few new highly connected papers. Report the rounds and thresholds (reproducibility).

Analyze (NetworkX)

import json, networkx as nx
g = nx.node_link_graph(json.load(open("kg/graph.json")), edges="links")
papers = [n for n, d in g.nodes(data=True) if d["type"] == "paper"]
cit = g.edge_subgraph([(u, v) for u, v, d in g.edges(data=True) if d["type"] == "cites"]).copy()

influential = sorted(cit.in_degree(), key=lambda x: -x[1])[:10]            # most cited within corpus
bridges = sorted(nx.betweenness_centrality(cit.to_undirected()).items(), key=lambda x: -x[1])[:10]
communities = nx.community.louvain_communities(cit.to_undirected(), seed=0)  # research clusters

Interpretation guidelines:

• High in-degree within the corpus = foundational for this literature (not globally important).
• High betweenness = bridges clusters, often the best papers to read for synthesis sections.
• Communities = candidate subsections of a review. Name each by its top topics and most cited paper.
• Sparse links between clusters that share topics = a potential research gap worth stating (and verifying).
• Co-authorship components show research groups. Beware name-based author merging without --openalex.

Visualize

• Gephi: open graph.graphml → ForceAtlas2 → size by in-degree → color by modularity class.
• Obsidian: open kg/vault as a vault and use Graph view, filtering by folder (papers/authors/topics).
• Small graphs in documents: convert the citation subgraph to Mermaid (flowchart LR, one edge per cites).

Integrity rules

• Never invent edges or papers. Every node and edge must come from the library or the OpenAlex record.
• Verify surprising structure, such as an apparent gap, by reading the papers before claiming it.

Caveats to state in any write-up

Coverage depends on OpenAlex/Crossref metadata; citation edges are only as complete as referenced_works. Keyword/topic nodes are coarse. Always report the data source, retrieval date and filtering thresholds.

Related skills

paper-corpus-rag (full-text QA over the same papers), literature-review, systematic-review-prisma, networkx, reference-manager-interop, citation-verification.