[Paper Review] Discovering Supply Chain Links with Augmented Intelligence

Achintya Gopal, Chunho Chang|arXiv (Cornell University)|Nov 2, 2021

Sentiment Analysis and Opinion Mining71 references4 citations

TL;DR

This paper proposes an augmented intelligence framework that combines graph neural networks (GNNs) with domain expertise to discover previously unknown supplier-customer relationships in global supply chains. By training a BERT-based GNN on company-level features and supply chain data, the model achieves high recall (up to 74% on customer-side predictions for Russia) and hit rates over 50% in key industries, significantly improving risk assessment in incomplete supply chain networks.

ABSTRACT

One of the key components in analyzing the risk of a company is understanding a company's supply chain. Supply chains are constantly disrupted, whether by tariffs, pandemics, severe weather, etc. In this paper, we tackle the problem of predicting previously unknown suppliers and customers of companies using graph neural networks (GNNs) and show strong performance in finding previously unknown connections by combining the predictions of our model and the domain expertise of supply chain analysts.

Motivation & Objective

• To address the challenge of incomplete and unreported supplier-customer relationships in global supply chains, which hinder accurate risk assessment.
• To improve the detection of previously unknown business connections by combining machine learning predictions with expert validation.
• To develop an inductive, generalizable model that can predict new edges for companies not seen during training.
• To evaluate model performance across diverse industries and geographies to ensure robustness and fairness in real-world deployment.
• To demonstrate that human-AI collaboration outperforms purely automated systems in identifying complex, nuanced supply chain relationships.

Proposed method

• Training a Graph Neural Network (GNN) on a global supply chain dataset with over 23,000 companies and 250,000 total connections, using company-level features such as industry, location, revenue segmentation, and financials.
• Employing BERT-large with mean pooling of the second-to-last layer embeddings to encode textual features like company descriptions and competitor lists, enhancing semantic understanding.
• Using an inductive learning approach that leverages external features per company rather than relying solely on node embeddings, enabling generalization to unseen companies.
• Applying bidirectional LSTMs to refine contextual representations of token embeddings, which improved performance over sentence-BERT baselines.
• Validating predictions by cross-referencing GNN outputs with primary sources (e.g., company filings, presentations) to confirm new edges.
• Constructing a stratified validation set to evaluate performance across industries and countries, accounting for data skew and ensuring reliable metrics.

Experimental results

Research questions

• RQ1 Can a GNN-based model effectively predict previously unknown supplier-customer relationships in supply chains with incomplete reporting?
• RQ2 How does incorporating domain-specific company-level features (e.g., industry, location, financials) improve link prediction performance compared to graph-only models?
• RQ3 To what extent does human-in-the-loop validation enhance the reliability and practical utility of AI-generated predictions in supply chain risk analysis?
• RQ4 How does model performance vary across different industries and countries, especially in data-scarce sectors like Financials and Real Estate?
• RQ5 What is the optimal combination of NLP and GNN components (e.g., BERT vs. RoBERTa, pooling strategies) for high-precision supply chain link discovery?

Key findings

• The model achieved a recall of 74% and hit rate of 59% for customer-side predictions from Russia, indicating strong performance in high-impact markets.
• Performance was highest in the Energy and Communications sectors, with customer-side recall exceeding 70% and hit rates above 50%.
• The Financials and Real Estate industries showed the lowest performance (recall ~37% and ~43% respectively), primarily due to sparse data and ambiguous relationship patterns.
• Using BERT-large with mean pooling of the second-to-last layer outperformed other architectures, including RoBERTa and sentence-BERT variants, especially when combined with bidirectional LSTMs.
• The model demonstrated strong generalization, achieving high performance on unseen companies due to its inductive learning design and rich feature integration.
• Stratified evaluation revealed that performance was reasonably uniform across countries, though low-observation buckets (e.g., one observation) should be interpreted with caution.