feat: citation network analysis and research cluster detection

[BenjaminNavet]

Description

Add a post-aggregation analysis module that builds a citation graph from the collected papers using existing CrossRef/OpenCitations data, and surfaces structural insights: research community clusters, highly-connected hub papers, and papers with recent citation surges. Output would include cluster labels per paper and optionally a graph file (DOT or JSON) for visualisation.

Justification

SciLEx already collects citation data but only uses raw counts for filtering. Treating those counts as a graph unlocks well-established bibliometric analyses (co-citation clustering, burst detection) that reveal research landscape structure without any new API calls.

Affected files

• new scilex/citation_network.py
• scilex/citations/citations_tools.py

[Mattheex]

I tested the idea on my side, and here's my conclusion.

OpenCitations
The SPARQL endpoint crashes at depth 2 in a single query. The workaround is splitting into 4 CONSTRUCT queries (one per direction × hop) and merging incrementally. However, you neither get title nor abstract with this approach.

doi.org + Crossref
The best approach for fetching citation data along with titles and authors is doi.org. With content negotiation, you can fetch everything in Turtle format directly. Combined with the Crossref API for reference lists, you get everything in one pass.

However, at depth 2 you get approximately 85k articles with only 1.3 edges per node on average. This is too sparse for meaningful clustering, and you mostly find papers related by their authors rather than by their research topics.

OpenAlex
For a scientific topic knowledge graph, OpenAlex would be ideal with its structured topic tags and abstracts, but it requires a paid plan.

Best Open Alternative
The best open alternative is to embed abstracts using thenlper/gte-large for similarity scoring, then pull topic labels from Wikidata. This is the same approach OpenAlex uses internally, it's completely free, and it avoids the author-bias problem of pure citation graphs.