We develop and deploy graph neural networks (GNN) for supply chain optimization – from demand forecasting to dynamic inventory management. When a disruption occurs in the graph of suppliers, plants, and warehouses, it cascades. Classic operations research methods (LP, MIP) lose structural patterns and generalize poorly to new topologies. GNN automatically extracts these patterns and adapts to network changes. Contact us to assess your case.
What remains unsolved with classical methods
Disruption propagation. When a tier-2 supplier fails, linear models miss the cascade effect through tier-1 to the final node. GNN sees it: disruption information propagates through the graph in O(depth) message-passing steps.
Demand forecasting in heterogeneous networks. SKUs in one region correlate with SKUs in another through shared seasonal and event factors. Standard univariate ARIMA or even SKU-level LightGBM does not capture these cross-node dependencies.
Dynamic routing. The optimal route depends on the current state of the entire network – warehouse loads, transport availability, delays. Heuristic routing does not adapt in real time.
How GNN models disruption propagation
The architecture builds on a heterogeneous graph: node types Supplier, Plant, DC (distribution center), Store/Customer; edges transports, supplies, distributes. We use Heterogeneous Graph Transformer (HGT) from PyTorch Geometric – it distinguishes node and edge types through type-specific attention heads. Each node has a feature vector: historical inventory levels, capacity, lead time, cost, geographic coordinates. Edges carry historical reliability, transport time, cost per unit.
Why STGNN outperforms classical forecasting methods
Formulation: predict demand at each node over a horizon of T steps, considering spatial dependencies (supply graph) and temporal dynamics. Architecture: STGNN = GCN for spatial dependencies + Temporal Convolution Network (TCN) or Transformer for time. Alternatively, Graph WaveNet (combines adaptive adjacency matrix with dilated TCN). On a real retail dataset (1,200 SKUs, 85 stores), Graph WaveNet achieved MAPE = 8.3% vs. 12.1% for independent SKU-level LightGBM. The improvement was especially noticeable during promotional periods with cross-store spillover effects. Savings on test projects reached 25% in operational costs.
Dynamic inventory management with RL+GNN
Inventory optimization in multi-echelon networks is a classic case where RL outperforms OR approaches in non-stationary environments. We formulate it as an MDP: state = current inventory levels at all nodes plus demand signals; action = order quantities along all edges; reward = -(stockout cost + holding cost + transportation cost). GNN-based policy network: state graph → GNN encoder → node/edge-level action head. Training via PPO in a supply chain simulator (custom environment based on OR-Gym or AnyLogic).
A challenge: reward shaping is critical. Without proper balance between stockout and overstock penalties, the agent quickly converges to a local minimum. A logarithmic penalty for overstock plus asymmetric stockout cost solves this better than linear weights. On a test environment (8-node, 3-echelon network, 52-week simulation), RL+GNN reduced total supply chain cost by 17% vs. a baseline (s,S)-policy while improving service level from 94.2% to 97.1%.Graph WaveNet paper
Disruption prediction
A separate task: predicting a specific supplier or transport leg delay/failure 2–4 weeks ahead. Features: external signals (weather API, port congestion indices, geopolitical risk scores via GDELT), internal data (historical on-time rates, lead time volatility). GNN aggregates signals from neighboring nodes: delays at a tier-2 supplier are early signals for tier-1. Binary classification achieved AUC-ROC of 0.79–0.83 on the task of predicting disruptions exceeding 5 days. Get a consultation on your case.
Tech stack
| Component | Tools |
|---|---|
| GNN | PyTorch Geometric, DGL |
| STGNN | Graph WaveNet, DCRNN |
| RL simulator | OR-Gym, custom Gym env |
| RL training | Stable-Baselines3, RLlib |
| OR solver | Google OR-Tools, Gurobi |
| Pipeline | Apache Airflow, dbt |
| Feature store | Feast, Redis |
Our process
| Stage | Duration | Outcome |
|---|---|---|
| Graph audit | 1–2 weeks | Inventory of nodes, edges, data sources |
| Feature engineering | 2–3 weeks | Historical time series, external signals |
| STGNN for forecasting | 3–4 weeks | Baseline MAPE, comparison with ARIMA/LightGBM |
| RL simulator | 2–3 weeks | Calibrate cost parameters to business logic |
| Shadow deployment | 2–4 weeks | Parallel run vs. current policy |
What's included in the work
- Graph audit of the current network and data sources.
- Feature engineering: preparation of time series and external signals.
- Building an STGNN demand forecasting model.
- Developing an RL simulator for multi-echelon inventory.
- Shadow deployment with A/B testing.
- Model and API documentation.
- Client team training (1–2 days).
- Support for 1 month after launch.
Our experience and guarantees
We have been doing ML-driven supply chain optimization for over 5 years, completing 15+ projects for retail and logistics. We guarantee a transparent model and integration into your existing infrastructure. Our specialists are certified in PyTorch and MLOps. We will assess your case in a free consultation – contact us, and we will select the optimal architecture for your budget and timeline.