Building the AI-Generated Production Network (AIPNET)
Our goal is to create a comprehensive map of how products are connected through production processes worldwide. This map, called the AI-generated Production Network (AIPNET), shows which products are used to make other products, helping us understand global trade and production better (see the full paper here).
Starting with Product Codes
We began with the Harmonized System (HS) product codes, an international standard that classifies over 5,000 products used in global trade. Each product has a unique code and a description, like “Full-fat milk and cream” or “Wind generators.”
The Challenge and Solution
While it’s easy to find information about how some products are connected (like milk being used to make cheese), doing this for thousands of products is challenging due to the vast amount of information.
To tackle this, we used advanced AI language models. These models have been trained on a wide range of internet text and can help us efficiently gather and organize information about product connections at a large scale.
Using AI to Map Product Connections
We developed a structured process to build AIPNET using AI tools. Here’s how we did it:
Step 1: Setting Up the AI Model
We used GPT-4o, a cutting-edge AI language model. We customized it for our task by designing specific prompts.
For each product code, we asked the AI model to list and describe other products that are directly connected in production—either products that are inputs to it or products that it helps to produce. We repeated this process multiple times for each product to get a comprehensive list of connections.
Step 2: Processing AI Output
The AI provided descriptions of related products in plain language. We analyzed these descriptions to identify the connections between products.
By combining the results from multiple rounds, we created an initial network where each product is a node, and the connections between them are edges.
Step 3: Matching Descriptions to Official Codes
To ensure accuracy, we needed to match the AI-generated product descriptions to the official HS codes.
We used a technique called text embeddings, which converts text into numerical representations that capture their meanings.
By comparing these numerical representations, we could find the closest matches between the AI descriptions and the official product codes, ensuring that our network reflects actual products.
Step 4: “Pruning” the Network
We wanted to make sure that each connection in our network represented a real production relationship.
We asked the AI model to verify each connection by confirming whether one product could realistically be used to make the other.
This step helped us eliminate any incorrect connections and improved the overall accuracy of AIPNET.
Understanding the Network
After building AIPNET, we analyzed its structure to understand how products are connected globally.
- Size and Scope: The network includes over 5,000 products and nearly 1 million connections between them.
- Connectivity: While the network isn’t densely connected overall, there are clusters where certain products are highly interconnected due to their roles in multiple production processes.
- Product Roles: We classified products based on their roles in the economy:
- Intermediate Goods: Products used to make other products (e.g., raw materials, components).
- Capital Goods: Products used to produce other goods or services (e.g., machinery).
- Final Consumption Goods: Products purchased by end consumers.
We found that intermediate and capital goods tend to have many connections because they are used in the production of numerous other products.
Validating AIPNET
To ensure that AIPNET accurately reflects real-world production relationships, we compared it to official data. We used Input-Output tables from the United States and Mexico, which show how industries use products from other industries.
Our findings showed a strong correlation between AIPNET and these official tables. This means that our AI-generated network successfully captures the actual connections in global production.
Measuring Product Importance: Integrated Global Product Centrality (IGPC)
We wanted to identify which products are most important in global trade, not just based on how much they’re traded but also on their role in production networks.
Introducing IGPC
The Integrated Global Product Centrality (IGPC) is a measure we developed that combines:
- Trade Data: How much a product is traded globally.
- Network Position: How connected a product is within AIPNET.
How IGPC Works
IGPC is inspired by algorithms like PageRank, which rank web pages based on their connections.
In our context, a product’s importance increases if it:
- Is heavily traded worldwide.
- Is a key input for many other products.
- Is connected to other important products.
By considering both trade volume and network connections, IGPC provides an integrated view of a product’s significance in the global economy.
In short
Through the use of advanced AI language models and careful validation, we’ve built AIPNET—a detailed map of how products are interconnected in global production.
This network helps us:
- Understand complex supply chains.
- Identify key products that are crucial for many industries.
- Analyze global trade patterns more effectively.
Our methodology showcases how AI can be leveraged to synthesize vast amounts of information, providing valuable insights into the interconnected nature of the global economy.
See the working paper for our main results.