Node Classification

The foundational mechanism where nodes iteratively aggregate feature vectors from local neighbors to update their own hidden state. In a Bill of Materials (BOM) Graph, a component node refines its representation based on the properties of its sub-components.

• GraphSAGE: Enables inductive learning by sampling a fixed-size neighborhood, crucial for dynamic supply chains with new suppliers.
• Graph Isomorphism Network (GIN): Achieves maximum discriminative power under the Weisfeiler-Lehman test, distinguishing subtle structural differences in multi-echelon networks.

Replaces uniform averaging with learnable importance weights, allowing the model to prioritize critical dependencies.

• Graph Attention Networks (GATs) assign dynamic weights to upstream suppliers, automatically identifying which Tier-2 vendors most influence a Tier-1 node's risk profile.
• Multi-head attention stabilizes learning by capturing different relational semantics simultaneously, such as separating cost dependencies from lead-time dependencies.

Real supply chains contain diverse node types (factories, SKUs, ports) and edge types (supplies, ships, stores). Relational Graph Convolutional Networks (R-GCNs) apply distinct transformation matrices per relation type.

• Prevents information blurring between distinct interaction modes.
• Enables simultaneous classification of supplier risk and part criticality within a single unified graph schema.

Integrates graph convolutions with recurrent units to classify nodes in evolving networks. Spatio-Temporal GNNs (ST-GNNs) model how a warehouse's congestion status propagates through the network over time.

• Captures the lag effect of a port closure rippling through downstream distribution centers.
• Essential for predicting dynamic lead-time classifications based on real-time traffic and weather telemetry.

Leverages unlabeled graph data to learn robust node representations before fine-tuning on scarce labeled examples. Graph Contrastive Learning creates augmented views of the supply chain via edge dropping or feature masking.

• Maximizes agreement between representations of the same factory under different stochastic perturbations.
• Dramatically improves classification accuracy when ground-truth disruption labels are rare.

Critical for operational trust. GNNExplainer identifies the minimal subgraph and feature mask responsible for classifying a node as 'high-risk'.

• Validates that the model flags a supplier due to a specific geopolitical subgraph, not a spurious correlation.
• Graph Robustness techniques defend against adversarial structural attacks where malicious actors manipulate edge data to hide supply chain vulnerabilities.

The core computational mechanism enabling node classification. In this process, a target node iteratively aggregates feature vectors from its immediate neighbors and updates its own hidden state. This allows label information to propagate across the graph structure. For a supplier node, this means its classification (e.g., 'at-risk') is influenced by the financial health signals of its connected manufacturers and raw material sources.

The process of mapping nodes into a low-dimensional, dense vector space where geometric proximity captures structural and feature similarity. These embeddings serve as the input for classification layers. Techniques range from shallow transductive methods (like Node2Vec) to deep inductive encoders (like GraphSAGE). A well-trained embedding places all 'bottleneck' warehouse nodes close together in the latent space, simplifying the final linear classification step.

The principle that connected nodes tend to share the same class label. Node classification performance is heavily dependent on this property. In a supply chain, high homophily exists if a delayed supplier is typically connected to other delayed suppliers. Low homophily (heterophily) requires specialized architectures that can learn edge-based distinctions rather than blindly smoothing features, as standard GCNs may fail when dissimilar nodes are frequently linked.

A critical capability for dynamic supply chains where new entities constantly appear. Unlike transductive methods that require the entire graph during training, inductive models like GraphSAGE learn a function to generate embeddings for previously unseen nodes. This allows a trained node classification system to instantly predict the category of a newly onboarded supplier based solely on its features and local neighborhood structure, without costly retraining.