Hybrid GNN for Supply Chain Demand Forecasting

A novel Graph Neural Network architecture combining local (GATv2) and global (TransformerConv) attention for next-day sales prediction in FMCG supply chains.

---

Overview

This project extends the SCG benchmark dataset (Wasi et al., 2025) with a custom hybrid GNN architecture and a temporally-aware data pipeline for supply chain demand forecasting. The model predicts next-day sales orders for 28 active FMCG products simultaneously using a graph of product relationships.

Key result: R² = 0.68, NRMSE = 0.071 — outperforming all baselines including LSTM, GAT-only, and Transformer-only models on the SCG dataset.

---

Architecture

The core contribution is a Hybrid GNN that runs two parallel attention paths and fuses them:

Input (28 features per node)
        │
  [Input Projection]  Linear(28 → 64)
        │
   ┌────┴────┐
   │         │
[GATv2]  [TransformerConv]       ← 2 layers each, 4 heads
Local     Global attention
   │         │
   └────┬────┘
   [Concatenate]  → (128,)
   [Fusion Layer] LayerNorm + GELU + Dropout
   [Output Head]  Linear → 1
        │
   Predicted sales (next day)

• GATv2 path — captures local neighbourhood effects (products sharing the same storage location)
• TransformerConv path — captures global patterns across the entire product graph
• Fusion — simple concatenation + linear projection learns which path matters more per context

---

Key Findings

1. Weekly temporal cycles dominate daily signals

Analysis of the SCG dataset revealed:

Lag-1 autocorrelation (today → tomorrow):     0.133  (weak)
Lag-7 autocorrelation (same day last week):   0.389  (3× stronger)

This motivated a 7-day sliding window input that improved R² from 0.50 → 0.68 (+36%).

2. Storage Location edges outperform Plant edges

Plant edges (paper default):      R² = 0.50
Storage Location edges (ours):    R² = 0.68

Products sharing storage locations have stronger demand correlations than products sharing manufacturing plants.

3. 12 of 40 products are near-zero and distort metrics

Products with >80% zero sales were identified and excluded from evaluation. Including them artificially inflates R² because predicting zero for always-zero products looks like a good prediction.

4. Graph structure consistently outperforms no-graph baseline

LSTM (no graph):      R² = 0.511
GAT Only:             R² = 0.573
Transformer Only:     R² = 0.581
Hybrid GNN (ours):    R² = 0.676  ← best

---

Results

Model	R²	RMSE (scaled)	NRMSE
Hybrid GNN (ours)	0.676	0.373	0.071
Transformer Only	0.581	0.424	0.081
GAT Only	0.573	0.429	0.082
LSTM (no graph)	0.511	0.458	0.088
Product Mean Baseline	0.472	—	—
Persistence (today=tomorrow)	0.128	—	—

---

Dataset

This project uses the SCG dataset from:

Wasi, A.T., Islam, M.S., Akib, A.R., & Bappy, M.M. (2025). Graph Neural Networks in Supply Chain Analytics and Optimization: Concepts, Perspectives, Dataset and Benchmarks. arXiv:2411.08550

• Source: Leading FMCG company in Bangladesh
• Period: January 1, 2023 – August 9, 2023 (221 days)
• Products: 40 SKUs (28 active after filtering)
• Features: Production, Sales Orders, Delivery to Distributors, Factory Issues
• Dataset repo: https://github.com/CIOL-SUST/SCG

---

Project Structure

supply-chain-gnn/
├── data/                          # Place SCG dataset files here
│   ├── NodesIndex.csv
│   ├── Edges (Storage Location).csv
│   ├── Edges (Plant).csv
│   ├── Edges (Product Group).csv
│   ├── Sales Order.csv
│   ├── Production .csv
│   ├── Factory Issue.csv
│   └── Delivery To distributor.csv
│
├── supply_chain_gnn.ipynb         # Main notebook (run top to bottom)
│
├── baselines.py                   # Standalone baseline training script
│
└── README.md

---

Installation

# 1. Clone the repository
git clone https://github.com/a34656/supply-chain-gnn.git
cd supply-chain-gnn

# 2. Install dependencies
pip install torch torchvision torchaudio
pip install torch-geometric
pip install pandas numpy scikit-learn matplotlib seaborn jupyter

Tested with:

• Python 3.9+
• PyTorch 2.0+
• PyTorch Geometric 2.3+
• CUDA 11.8 (optional, CPU also works)

---

Usage

Option A — Jupyter Notebook (recommended)

jupyter notebook supply_chain_gnn.ipynb

Update DATA_DIR in Cell 2 to point to your data folder:

DATA_DIR = r'path/to/your/data'

Then run all cells top to bottom. Cell order matters:

Cell 1  → Imports
Cell 2  → Data pipeline (edges, features, 7-day window)
Cell 3  → Scaling (train-only fit, no leakage)
Cell 4  → Build graph snapshots
Cell 5  → Model definitions
Cell 6  → Training helpers
Cell 7  → Train all 4 models
Cell 8  → Results table
Cell 9  → Visualizations
Cell 10 → Save predictions

Option B — Baselines script

python baselines.py

---

Data Pipeline

221 days × 40 products × 4 features
        ↓
Remove 12 dead products (>80% zero sales)
        ↓
221 days × 28 products × 4 features
        ↓
7-day sliding window
        ↓
213 samples × 28 products × 28 features  (7 days × 4 features)
        ↓
80/20 temporal split
        ↓
Train: 170 snapshots  |  Test: 43 snapshots

Important: The train/test split is strictly temporal — the model is always trained on earlier data and tested on later data. No shuffling.

---

Graph Construction

Nodes are products. Edges connect products sharing the same storage location (3,046 edges across 28 active nodes). Self-loops are added so each node aggregates its own features. Edge weights are degree-normalized.

# Storage Location edges — best performing edge type
edges_df = pd.read_csv('Edges (Storage Location).csv')

# Self-loops added for all nodes
from torch_geometric.utils import add_self_loops
ei_tensor, ew_tensor = add_self_loops(ei_tensor, ew_tensor,
                                       fill_value=1.0, num_nodes=N_NODES)

Edge type comparison:

Edge Type	Edges	Isolated Nodes	R²
Storage Location	3,046	0	0.68
Plant	1,647	5	0.50
Product Group	188	0	not tested

---

Model Configuration

# Hybrid GNN — best model
HybridGraphModel(
    in_channels    = 28,    # 7 days × 4 features
    hidden_channels= 64,
    num_layers     = 2,
    dropout        = 0.4
)

# Training
optimizer = AdamW(lr=3e-4, weight_decay=1e-5)
scheduler = ReduceLROnPlateau(patience=10)
max_epochs = 200
patience   = 40   # early stopping

---

Comparison with Original Paper

The SCG paper (Wasi et al., 2025) benchmarks standard temporal GNNs with default hyperparameters and Plant edges. This work extends it with:

	SCG Paper	This Work
Architecture	DCRNN, TGCN, GConvGRU (off-the-shelf)	Custom Hybrid GNN (novel)
Input	Single-day snapshot (4 features)	7-day window (28 features)
Edge type	Plant edges only	Storage Location (empirically chosen)
Dead product analysis	Not addressed	12 products identified and excluded
GNN improvement over LSTM	~9% RMSE reduction	~18% RMSE reduction

---

Citation

If you use this code, please cite the original dataset paper:

@article{wasi2025scg,
  title={Graph Neural Networks in Supply Chain Analytics and Optimization:
         Concepts, Perspectives, Dataset and Benchmarks},
  author={Wasi, Azmine Toushik and Islam, MD Shafikul and
          Akib, Adipto Raihan and Bappy, Mahathir Mohammad},
  journal={arXiv preprint arXiv:2411.08550},
  year={2025}
}

---

License

This project is released under the MIT License. The SCG dataset is available under CC BY 4.0 at DOI: 10.5281/zenodo.13652826.

---

Acknowledgements

• SCG dataset by Wasi et al. (SUST & LSU)
• PyTorch Geometric for GNN implementations
• GATv2: Brody et al. (2022), TransformerConv: Shi et al. (2021)