Skill B: Network Analysis

This is the second Skill for the Network-aware Daily arXiv Research Briefing Agent. It consumes paper metadata from Skill A and produces research-network analysis for daily briefings.

• Builds a coauthorship graph from paper authors.
• Builds a paper-similarity graph from embeddings or a TF-IDF fallback.
• Runs Louvain and Label Propagation community detection.
• Computes PageRank, betweenness, and degree centrality.
• Identifies bridge authors with betweenness * (1 - clustering_coefficient).
• Exports JSON plus static PNG visualizations.

python -m pip install -e ".[dev]"

The core implementation uses only NetworkX and Matplotlib. PyVis is optional:

python -m pip install -e ".[interactive]"

skill-b-network analyze --input examples/papers.json --output outputs/network_analysis.json

Local module form:

python -m skill_b_network.cli analyze --input examples/papers.json --output outputs/network_analysis.json

The command writes:

• outputs/network_analysis.json
• outputs/artifacts/coauthorship_network.png
• outputs/artifacts/paper_similarity_network.png
• outputs/artifacts/community_method_comparison.png
• outputs/artifacts/interactive_network.html if PyVis is installed

Required top-level fields:

• run_id: run identifier used in reports.
• papers_metadata[]: paper records.

Required paper fields:

• paper_id
• arxiv_id
• title
• abstract
• authors
• categories
• published_at

Optional fields:

• embeddings: {paper_id: [float, ...]}.
• history_snapshots: previous-day snapshots for future dynamic metrics.
• graph_params: overrides for defaults.

Default graph params:

{
  "k_neighbors": 5,
  "sim_threshold": 0.55,
  "coauthor_weight": "count",
  "community_methods": ["louvain", "label_propagation"],
  "random_seed": 42,
  "top_n": 10,
  "visualization": {
    "max_nodes": 60,
    "label_top_n": 30,
    "depth": 1,
    "seed_nodes": 8,
    "min_component_size": 1
  }
}

Visualization params affect the PNG/HTML artifacts only; the analysis JSON still contains the full graph.

• max_nodes: maximum nodes drawn in one static network image.
• label_top_n: highest-ranked nodes that receive text labels.
• depth: neighbor expansion depth from the highest-ranked seed nodes when the full graph is too large.
• seed_nodes: number of central seed nodes used before depth expansion.
• min_component_size: hides tiny disconnected components in images when set above 1.

The output JSON contains:

• graphs: graph statistics for coauthorship and paper-similarity graphs.
• coauthor_edges: weighted author-author edges.
• paper_sim_edges: weighted paper-paper similarity edges.
• communities: method-specific assignments, community counts, and modularity.
• centrality: PageRank, betweenness, and degree centrality values.
• top_influencers: top authors and papers by PageRank.
• top_bridges: top bridge authors with explainability fields.
• emerging_communities: paper community summaries for the daily briefing.
• network_stats: repeated graph stats for downstream compatibility.
• artifacts: generated figure paths.
• warnings: recoverable data-quality and fallback messages.

python -m pytest

Skill A should pass its ranked paper metadata and embeddings to this Skill. If Skill A has not implemented embeddings yet, this Skill still runs by using TF-IDF cosine similarity over title + abstract.

Skill B does not fetch arXiv data and does not summarize papers. It only analyzes metadata and vectors provided by the upstream retrieval/summarization layer.

• Author disambiguation is simple normalized-name matching.
• Louvain and Label Propagation are internal community estimates, not ground truth.
• Dynamic metrics require historical snapshots; without them, growth_7d, novelty, and bridge_pressure are null.