Heterogeneous Graph

Unlike homogeneous graphs, heterogeneous graphs contain distinct node types (e.g., suppliers, warehouses, products) and distinct edge types (e.g., 'supplies', 'stores', 'transports'). Each type can carry its own feature schema. For example, a 'warehouse' node may have attributes like capacity and location, while a 'product' node has SKU and weight. This typed structure allows models to learn type-specific transformations.

R-GCNs are the foundational neural architecture for heterogeneous graphs. They assign a distinct weight matrix per relation type during neighbor aggregation. The update rule for a node is:

• h_i = σ( Σ_r Σ_j∈N_i^r (1/c_i,r) * W_r * h_j + W_0 * h_i ) This allows the model to learn that the 'supplies' relationship has a different semantic meaning than the 'competes_with' relationship, preventing information conflation.

Metapaths define composite semantic relationships across heterogeneous graphs. A metapath like Supplier → (supplies) → Warehouse → (stores) → Product captures the indirect relationship between a supplier and the products ultimately stored. Algorithms like metapath2vec perform random walks constrained by these patterns to generate contextually rich node embeddings that preserve the graph's heterogeneous structure.

A heterogeneous graph is governed by a schema that formally defines the permitted node and edge type combinations. This schema acts as a constraint and a prior for learning. For instance, an edge type 'manufactures' may only be permitted between nodes of type 'Factory' and 'Product'. Aligning this schema with domain ontologies ensures the graph's structure reflects real-world business logic and prevents nonsensical connections.

HAN introduces a dual-level attention mechanism for heterogeneous graphs:

• Node-level attention: Learns the importance of different neighboring nodes connected by the same metapath.
• Semantic-level attention: Learns the importance of different metapaths themselves. This allows the model to dynamically determine that the metapath Supplier → Part → Product is more critical for predicting supply risk than Supplier → Region → Supplier.

Heterogeneous graphs are the ideal data structure for supply chain digital twins. They can natively model:

• Physical entities: Factories, ports, trucks, inventory lots.
• Commercial entities: Purchase orders, invoices, contracts.
• External factors: Weather events, geopolitical zones, traffic grids. The diverse edge types (e.g., 'delayed_by', 'fulfills', 'impacts') allow for complex 'what-if' simulations and root cause analysis across the entire value chain.

A specialized graph neural network architecture explicitly designed for heterogeneous graphs. Unlike standard GCNs that treat all edges identically, R-GCNs apply distinct weight matrices for each relation type during neighbor aggregation. This allows the model to learn how a 'supplies' relationship differs from a 'transports' relationship when updating node embeddings. In a supply chain context, an R-GCN can simultaneously process supplier-part relationships, warehouse-inventory links, and carrier-lane assignments within a single unified model, preserving the semantic meaning of each connection type.

A technique for mapping the entities and relations of a knowledge graph into a continuous, low-dimensional vector space. The goal is to preserve the graph's structural and relational information so that mathematical operations in the embedding space reflect real-world semantics. For heterogeneous supply chain graphs, embedding methods like TransE or RotatE can encode complex relational patterns such as 'supplier A provides component B to manufacturer C.' These embeddings serve as feature inputs for downstream node classification and link prediction tasks.

A low-dimensional vector representation of a node that encodes both its intrinsic features and its structural position within a graph. In a heterogeneous graph, node embeddings must capture not just proximity but also the diversity of relationship types connecting entities. For example, a supplier node's embedding would simultaneously encode its material categories, geographic location, and the reliability scores of its connections to downstream manufacturers. These embeddings are the primary input for machine learning tasks such as node classification and anomaly detection.

The foundational computational mechanism of graph neural networks. During each iteration, nodes aggregate feature information from their immediate neighbors, transform it via a learned function, and update their own representation. In a heterogeneous graph, message passing must account for multiple edge types by using relation-specific transformations. A manufacturer node might simultaneously receive messages from upstream suppliers (material availability signals) and downstream distributors (demand forecast signals), with each message type processed through distinct neural pathways before being integrated into the node's updated state.

The task of predicting the existence or likelihood of a missing or future connection between two nodes. In a heterogeneous supply chain graph, link prediction can forecast potential supplier-buyer relationships, identify alternative sourcing options, or detect missing bill-of-materials dependencies. Models leverage the learned node embeddings and known relation patterns to score candidate triples. A high score for a 'can_fulfill' edge between a new supplier and a manufacturer might trigger an automated sourcing recommendation.

A GNN variant that employs self-attention mechanisms to assign different importance weights to neighboring nodes during aggregation. When applied to heterogeneous graphs, attention allows the model to dynamically learn which relationship types and which specific neighbors are most relevant for a given task. For instance, when predicting disruption risk for a factory, the attention mechanism might learn to weigh tier-1 supplier status more heavily than tier-3 logistics connections, automatically prioritizing the most critical dependencies without manual feature engineering.