From Connections to Meaning: Why Heterogeneous Graph Transformers (HGT) Change Demand Forecasting

Standard demand forecasting models often fail by ignoring structural relationships between products, and while simple Graph Neural Networks (GraphSAGE) help, they treat all connections equally. This article introduces Heterogeneous Graph Transformers (HGT) to create relationship-aware forecasts that understand the difference between connections like a shared plant versus a shared product group. By learning the specific meaning of each relationship, HGT can more accurately model how demand shocks propagate through a supply chain network. An implementation on a real-world dataset shows HGT reduces forecast error by an additional 6-7% over GraphSAGE, leading to fewer misallocated units and more stable operations. This turns a topologically connected forecast into a more meaningful, operationally grounded prediction.