Multi-Echelon Graph

The defining characteristic is the explicit modeling of echelons—distinct, sequential stages in the supply chain. Nodes are organized into tiers (Tier 1, Tier 2, Tier n suppliers) with directed edges representing the flow of goods from raw material sources to end consumers. This structure enables bullwhip effect analysis by isolating how demand signal distortion amplifies as it propagates upstream through successive echelons.

Multi-echelon graphs are typically directed acyclic graphs where material flows in one direction (upstream to downstream) and information/ financial flows in the reverse. This acyclic constraint prevents circular dependencies that would violate physical supply chain logic. The DAG structure enables efficient topological sorting for critical path analysis and lead time accumulation across the entire network.

These graphs are inherently heterogeneous, containing multiple node types (suppliers, factories, cross-docks, regional warehouses, last-mile hubs, retail stores) and edge types (supplies, ships-to, fulfills, returns). Each node type carries distinct feature vectors—suppliers have capacity and lead time attributes, while warehouses have storage constraints and throughput rates. This heterogeneity enables relational graph convolutions that apply separate weight matrices per relationship type.

Edges encode multi-commodity flow constraints representing the simultaneous movement of different SKUs, raw materials, and sub-assemblies through shared logistics channels. Edge features include capacity limits, transportation modes, tariffs, and carbon emissions per unit. This allows the graph to model bundled logistics where multiple products share container space or truckload capacity across echelons.

Multi-echelon graphs often embed a Bill of Materials subgraph at manufacturing nodes, representing the hierarchical assembly structure where parent products depend on child components. This creates a dual-layered graph: the macro-level logistics network overlaid with micro-level production recipes. The BOM explosion across echelons enables precise material requirements planning (MRP) and component-level disruption impact analysis.

Edge weights represent stochastic lead times—the probabilistic delay between order placement at one echelon and receipt at the next. These temporal properties transform the static graph into a spatio-temporal graph where node states (inventory levels, backorders) evolve based on upstream delays. Graph neural networks can learn to propagate delay signals through the topology, enabling predictive lead time analytics that account for cascading tier-n disruptions.

The core computational mechanism where nodes iteratively aggregate feature information from their neighbors to update their own representations. In a multi-echelon context, a warehouse node receives demand signals from downstream retailers and inventory status from upstream suppliers simultaneously.

• Aggregation function: Sum, mean, or max pooling of neighbor states
• Update function: Combines aggregated neighbor data with the node's current state
• Hop count: Determines how far information propagates across echelons

Employs self-attention mechanisms to assign different importance weights to neighboring nodes during feature aggregation. This is critical for supply chains where a disruption at a Tier-1 supplier should receive higher attention than a stable Tier-3 node.

• Learns implicit attention coefficients per edge
• Enables the model to focus on bottleneck suppliers
• Multi-head attention captures diverse relationship patterns

A graph structure containing multiple types of nodes and edges, representing diverse entity and relationship categories within a single network. A multi-echelon graph is inherently heterogeneous, with nodes like Supplier, Warehouse, Distribution Center, and Retailer connected by edges like ships_to, supplies, and fulfills.

• Requires type-specific weight matrices
• Captures semantic meaning of different relationships
• Enables relational reasoning across echelons

Designed specifically for heterogeneous graphs, R-GCNs apply distinct weight matrices for different relation types during neighbor aggregation. For a multi-echelon graph, the model learns separate transformations for 'supplier-to-warehouse' edges versus 'warehouse-to-retailer' edges.

• Handles multi-relational data natively
• Prevents information leakage between relation types
• Basis-decomposition variant reduces parameter explosion

Models dynamic systems by simultaneously capturing spatial dependencies via graph convolutions and temporal dependencies via recurrent or convolutional units. Essential for multi-echelon graphs where inventory levels and shipment statuses evolve over time across the network topology.

• Spatial component: GCN on the supply chain topology
• Temporal component: LSTM or TCN on historical sequences
• Predicts future stockouts and delay cascades

The task of predicting the existence or likelihood of a missing connection between two nodes. In multi-echelon graphs, link prediction identifies potential alternative suppliers, hidden dependencies, or emergent logistics pathways that are not explicitly recorded in the ERP system.

• Uses node embeddings and edge scoring functions
• Discovers shadow connections in the supply network
• Enables proactive supplier diversification strategies