Methodology in Detail

Overview

Using the EventRegistry API, we collected 3,000 news articles related to the aerospace and defense industry. EventRegistry is a service that aggregates articles from more than 100 international media sources and provides them as structured data.

Query Specification

• Category: dmoz/Business/Aerospace_and_Defense
• Article count: 3,000
• Language: English
• Period: Most recent articles from 2024
• Sources: Bloomberg, Reuters, The Guardian, Defense News, Aviation Week, and others

Retrieved Data Fields

• Title (title)
• Body (body)
• Source name (source)
• URL (url)
• Publication timestamp (publishedAt)

Execution Time

The API request completed in about 1.5 minutes and was saved as raw_articles.json.

Parallel Processing Architecture

To process 3,000 articles efficiently, we built a parallel processing system using ThreadPoolExecutor. 100 workers run concurrently, each calling the Gemini 2.5 Flash Lite API to extract entities and relationships.

Prompt Design

We instructed the Gemini API to extract the following information:

Entities

• entity_name: The formal name of the entity
• entity_type: company, government_organization, government, subsidiary, region, university, research_institute
• entity_role: manufacturer, supplier, customer, regulator, research, and others
• country_code: ISO 3166-1 alpha-3 code (USA, CHN, JPN, etc.)

Relationships

• entity_from: The origin entity of the relationship
• entity_to: The destination entity of the relationship
• relationship_type: supplies_component, delivers_product, partners_with, acquires, invests_in, and others
• relationship_status: active, planned, completed, cancelled
• component_category: engine, avionics, materials, software, and others
• description: A detailed description of the relationship

Output Format

We required Gemini to output in structured JSON format, following the schema below:

{
  "entities": [
    {
      "entity_name": "Boeing",
      "entity_type": "company",
      "entity_role": "manufacturer",
      "country_code": "USA"
    }
  ],
  "relationships": [
    {
      "entity_from": "Spirit AeroSystems",
      "entity_to": "Boeing",
      "relationship_type": "supplies_component",
      "relationship_status": "active",
      "component_category": "fuselage",
      "description": "Spirit AeroSystems supplies fuselage sections for Boeing 737 MAX"
    }
  ]
}

Processing Results

• Success rate: 98.6% (2,959/3,000 articles)
• Entities extracted: 5,170 (before deduplication)
• Relationships extracted: 14,684 (before deduplication)
• Execution time: about 10 minutes (100 workers in parallel)

The Problem

Among the extracted entities, the same entity was often extracted multiple times due to naming variants and abbreviations. Example: "NASA," "National Aeronautics and Space Administration," "US Space Agency."

Solution Approach

Using semantic similarity judgments with Gemini 2.5 Flash Lite, we identified duplicates according to the following criteria:

• Matches between formal names and abbreviations (e.g., NASA ↔ National Aeronautics and Space Administration)
• Explicit subsidiary–parent relationships (e.g., Spirit AeroSystems ↔ Boeing Spirit)
• Regional naming variants (e.g., United States ↔ USA ↔ US)
• Minor differences in notation (e.g., Lockheed Martin ↔ Lockheed Martin Corporation)

Deduplication Process

1. Present all entities to Gemini
2. Generate groups of entities judged to be semantically identical
3. Select a representative entity from each group (the most frequently occurring name)
4. Unify the entity IDs in the relationship data to the representative entity

Results

• Entities: 5,170 → 4,620 (10.6% reduction)
• Relationships: 14,684 → 14,106 (3.9% reduction, duplicate relationships removed)

Graph Construction

We constructed a directed graph with 4,620 nodes and 8,134 edges. Nodes represent entities and edges represent relationships. Each node was assigned attributes such as entity_type, country, and role.

Centrality Analysis

We computed the following centrality measures:

Degree Centrality

A measure based on the number of connections a node has. Nodes with high degree centrality are hub entities that maintain many trade relationships.

• Top 1: SpaceX (245 connections)
• Top 2: NASA (236 connections)
• Top 3: United States (213 connections)

Betweenness Centrality

Measures how often a node lies on the shortest paths between other nodes. Nodes with high betweenness centrality act as "bridges" in the supply chain and pose a Single Point of Failure (SPOF) risk.

PageRank

Based on the same principle as Google's search algorithm, this measure gives higher weight to nodes that are connected to from important nodes. It measures "influence" within the aerospace and defense industry.

• Top 1: NASA (PageRank 0.0133)
• Top 2: United States (0.0109)
• Top 3: SpaceX (0.0091)

Inter-Country Flow Analysis

Based on the nationality of each entity, we aggregated trade flows between countries. This makes it possible to visualize which countries supply components and technology to which other countries.

Major Inter-Country Flows

• USA → FRA: 189 relationships (e.g., Boeing–Safran and GE–Safran joint engine development)
• USA → USA: 1,245 relationships (domestic supply chain)
• USA → GBR: 142 relationships (cooperation with Rolls-Royce)

Supplier Diversity Analysis

For each entity, we computed the following measures:

• Number of suppliers: mean 2.81 (median 2)
• Number of supplier countries: mean 1.55 (median 1)

We found that the vast majority of entities source from only 1–2 countries, making them vulnerable to geopolitical risk.

Connected Components Analysis

• Number of weakly connected components: 553 (independent clusters)
• Largest connected component size: 3,733 nodes (80.8% of the whole)
• Network density: 0.000381 (a very sparse network)

Force-Directed Graph

Using the force simulation of D3.js v7, we simulated the attraction and repulsion between nodes to generate a natural layout.

Logarithmic-Scale Node Sizing

To solve the problem of gigantic hub nodes (245 connections) obscuring other nodes, we computed node radii on a logarithmic scale:

radius = 5 + Math.log(connections + 1) * 5

This keeps radii within a minimum of 5px and a maximum of about 50px, greatly improving legibility.

Visual Hierarchy

• Opacity: The more connections, the more opaque (0.7–1.0)
• Stroke width: More important nodes have thicker borders (1–3px)
• Drop shadow: A shadow is added to nodes with 20 or more connections

Interactive Features

• Entity type filter: Narrow by company, government_organization, subsidiary, etc.
• Country filter: Choose from 20 countries (USA, CHN, IND, GBR, FRA, etc.)
• Relationship type filter: supplies_component, partners_with, etc.
• Top N node display: Show only the top N nodes by connection count
• Mouseover tooltip: Displays node name, type, role, country, and connection count
• Edge click: Displays detailed information about the relationship

Theoretical Foundations

In the policy analysis, we integrated the following theoretical frameworks:

Edler & Georghiou Policy Typology (2007, 2012)

• Supply-side policy: R&D subsidies, technology infrastructure development
• Demand-side policy: Public procurement, demand creation through regulation
• Systemic policy: Network building, standardization, ecosystem design

Mazzucato's Entrepreneurial State (2013, 2024)

• Government as an actor that takes risks and creates new markets, not merely a complement to the market
• Mission-oriented policy: an integrated approach directed at clear social goals (moon landing, climate change countermeasures)
• Examples: NASA Commercial Crew Program (cultivating SpaceX), DARPA (development of the internet)

Barabási-Albert Scale-Free Network Theory (1999)

• Preferential attachment: nodes that already have many connections are more likely to acquire new connections
• Power-law distribution: P(k) ~ k^(-γ), a few hubs and many peripheral nodes
• Network robustness and fragility: strong against random failures but weak against targeted attacks

Industrial Organization Theory

• Barriers to entry: commercial aircraft development costs of $5-20B (Boeing 787 was $32B)
• Natural monopoly: economies of scale, learning curves, network effects
• Lock-in effects: sunk costs, complementary assets, coordination costs

Institutional Analysis

We identified the key institutional factors that shape the network structure:

ITAR (International Traffic in Arms Regulations)

• If even a single US-made component is included, the entire product becomes subject to regulation (cascade effect)
• 2024 AUKUS reform: 70% of items became license-free among Australia, the UK, and the US

India's Offset Policy

• Contracts exceeding 300 crore INR (about $36M) carry a 30% domestic production obligation
• Effect: defense market size grew from $27.1B to $54.4B (2014-2024)

NATO Standardization Agreements (STANAG)

• De facto adoption of US standards: interoperability requirements give US products an advantage

Verification Procedures

• Entity verification: Manually confirming that major companies (Boeing, Airbus, SpaceX, etc.) were correctly extracted
• Relationship verification: Verifying that known trade relationships (e.g., the GE-Safran engine joint venture) were correctly recorded
• Country code verification: Confirming compliance with the ISO 3166-1 alpha-3 standard
• Deduplication verification: Sampling to confirm that semantically identical entities were merged

Acknowledging Data Limitations

• Bias: English-language media only, with a US- and Europe-centric reporting tendency
• Omissions: Military secrets and non-public contracts are not included
• Temporal constraint: Only 2024 articles; the past evolution of the network is not analyzed
• Industry intermixing: Aircraft, space, and defense are intermixed, so simple comparisons are inappropriate

Available Data

• entities_deduplicated.csv (4,620 entities)
• relationships_deduplicated.csv (14,106 relationships)
• network_statistics.json (overall network statistics)
• critical_entities.csv (top 100 entities by centrality)
• country_flow_matrix.csv (inter-country flow matrix)

Code

• src/news_fetcher.py (EventRegistry API data retrieval)
• extract_to_out_parallel.py (Gemini parallel extraction)
• deduplicate_to_out.py (semantic deduplication)
• analyze_policy.py (NetworkX analysis)
• visualize_network_d3.html (D3.js visualization)

Time-Series Analysis

Collect multi-year data to analyze the evolution of the supply chain network (e.g., the impact of COVID-19 and the war in Ukraine).

Community Detection

Automatically detect industry clusters (space, defense, commercial aircraft) using algorithms such as the Louvain method and Label Propagation.

Simulation

Simulate the impact of removing specific nodes (corporate bankruptcy, trade sanctions) on the network as a whole.

Machine Learning

Use graph neural networks (GNNs) for link prediction (forecasting future trade relationships) and anomaly detection (unnatural trade patterns).