Literature Citation Network Analyzer

Pull a paper's citation neighborhood from the public Semantic Scholar API (no key required), build a NetworkX directed citation graph, compute PageRank/community structure, view it as an interactive Plotly network.

Why this exists

"Who cites this paper, and what does it cite?" is a graph question, not a list question — PageRank identifies locally-influential papers, community detection surfaces thematic clusters, and the network layout itself shows structure a flat reference list can't.

Data

Default seed is the author's own published paper — Sakata et al., Radiative cooling film enabled by droplet-like infrared hot spots via low-cost and scalable spray-coating process for tropical regions, Cell Reports Physical Science 2024. Real, public, verifiable data (same as what's on the author's Google Scholar profile) — any DOI works via the sidebar input.

No API key required or used. Semantic Scholar's public keyless tier is rate-limited; fetch.py retries with exponential backoff on 429s and caches every response locally (cache/, gitignored) so repeated runs and Streamlit's rerun-on-every-interaction model don't hammer the API. A failed fetch raises S2UnavailableError rather than silently returning an empty network — an API outage is not the same fact as "this paper has no citations."

Pipeline

• fetch.py — Semantic Scholar API client: paper lookup, citations, references, retry/backoff, local JSON cache
• graph.py — pure functions (no network access, fully unit-testable): build a networkx.DiGraph from fetched records, compute PageRank/degree metrics, detect communities via greedy modularity
• app.py — Streamlit UI: DOI input, interactive Plotly network (node size = PageRank, color = community, ★ = seed paper), sortable table of most-central papers

Tests

tests/test_graph.py uses hand-built fixture data (no live API calls — offline, fast, deterministic):

• Edge direction correctness (citing → seed, seed → cited — easy to get backwards, tested explicitly)
• No duplicate nodes for a repeated paper
• PageRank sums to exactly 1.0
• In-degree matches the actual citation edge count
• No self-loops
• Every node lands in exactly one community
• Single-node graph doesn't crash metrics or community detection

Run it

pip install -r requirements.txt
pytest tests/          # 8 tests, all offline
streamlit run app.py

Project structure

citation_network_analyzer/
├── fetch.py                 # Semantic Scholar API client + cache
├── graph.py                 # NetworkX graph build + metrics (pure, testable)
├── app.py                   # Streamlit UI
└── tests/test_graph.py      # offline fixture-based graph-correctness tests