ChainSight-AI — Network-Aware Supply Chain Intelligence

ChainSight-AI is a Graph Neural Network (GNN) and Social Network Analysis (SNA) framework designed to model supply chain ecosystems as heterogeneous graphs. By combining time-series node metrics with multi-relational graph attention mechanisms, ChainSight-AI predicts product demand, forecasts factory issue events, and automatically pinpoints critical supply chain bottlenecks ("weakest links").

---

📌 Features

• Heterogeneous Graph Representation: Models 41 products connected across 4 distinct supply chain relationship layers (by_plant, by_storage, by_group, by_sub_group).
• Temporal GAT Architecture: Fuses a GRU temporal encoder with multi-head HeteroConv Graph Attention Networks to capture spatiotemporal dependencies.
• Dual Prediction Tasks:
  • Regression: Next-day sales order weight forecasting.
  • Classification: Next-day factory issue event prediction (binary).
• Risk Propagation & Bottleneck Extraction: Extracts multi-head attention weights to rank critical inter-product dependencies and visualize supply chain risk maps.
• Benchmark Baseline: Includes fully-scaled Multi-Layer Perceptron (MLP) baselines demonstrating GAT superiority.

---

🛠 Project Structure

SNAProject/
├── Raw Dataset/               # Raw supply chain nodes, edges, and temporal metrics
│   ├── Edges/                 # Edges across plants, groups, sub-groups, storage
│   ├── Nodes/                 # Product definitions and group mappings
│   └── Temporal Data/         # 221 days of daily unit & weight time-series
├── PreProcessing/             # Pipeline scripts for dataset preparation
│   ├── preprocess.py          # Assembles feature tensors and edge arrays
│   ├── mismatch.py            # Column consistency diagnostics
│   ├── hetero.py              # PyG HeteroData structure builder
│   ├── CreateSamples.ipynb    # Sliding-window sample generator (7-day history → 1-day forecast)
│   └── traintestsplit.ipynb   # Chronological 70/15/15 train/val/test splitter
├── GATCls.ipynb               # Temporal GAT Classification & Attention Weight Extractor
├── GATReg.ipynb               # Temporal GAT Regression
├── MLPCls.ipynb               # Baseline MLP Classifier
├── MLPReg.ipynb               # Baseline MLP Regressor
├── RiskImportanceGraph.ipynb  # Supply chain risk map visualization
├── GAT_Risk_Importance_Map.csv# Extracted edge importance rankings
└── Risk_Propagation_Map.png   # Rendered risk network visualization

---

📊 Results & Performance

Task	Metric	Baseline MLP	ChainSight GAT	Improvement
Classification (Factory Issue)	F1-Score	0.8663	0.9049	+4.46%
Classification (Factory Issue)	Accuracy	0.8574	0.8891	+3.70%
Regression (Sales Order Weight)	RMSE	18.2090	16.2003	-11.03% (Lower error)
Regression (Sales Order Weight)	MSE	331.5661	262.4504	-20.85% (Lower error)

---

🚀 Getting Started

Prerequisites

• Python 3.10+
• PyTorch
• PyTorch Geometric (torch_geometric)
• scikit-learn, pandas, numpy, matplotlib, networkx

Quickstart Workflow

1. Preprocess raw data:

   python PreProcessing/preprocess.py

2. Generate sliding-window samples & split datasets: Run PreProcessing/CreateSamples.ipynb and PreProcessing/traintestsplit.ipynb.

3. Train Temporal GAT Model: Execute GATCls.ipynb or GATReg.ipynb.

4. Extract & Visualize Supply Chain Risk Map: Run RiskImportanceGraph.ipynb to generate Risk_Propagation_Map.png.