Enhancement of Graph Convolutional Networks Applied in Dynamic Rice Supply Chain Vulnerability Prediction Under Climate and Demand Shocks

Undergraduate Thesis in Computer Science
Author: Daniel Hardy C. Camacho
Degree: Bachelor of Science in Computer Science (BSCS)
Institution: Department of Computer Science, College of Information Systems and Technology Management, Pamantasan ng Lungsod ng Maynila (University of the City of Manila)

---

Source of truth

The canonical reference for this project is documentation/[RW] - GCN in Rice Ch. 1-5 UPDATED.pdf. Every reported table, metric, and result must match that PDF. Where any other document, extracted text, guide, or code comment disagrees with it, the PDF wins. The slide deck documentation/[GCN] - Rice Vulnerability Prediction - PPT.pdf is a defense presentation artifact and is not authoritative.

---

Research Overview

Agricultural supply chains operate as multi-tiered, time-varying networks subject to compounding disruptions such as climate extremes and demand shocks. Conventional Graph Convolutional Networks (GCNs) suffer from critical architectural constraints when applied to dynamic supply chains:

1. Static Topology Assumption: Baseline GCNs rely on a fixed adjacency matrix, failing to capture dynamic logistical rerouting and link degradation over time.
2. Closed-System Feature Initialization: Standard node features are treated as static closed systems, unable to ingest exogenous climate and market perturbations after initialization.
3. Deep Propagation Degradation: Directional agnosticism (ignoring forward supply vs. reverse demand flows), over-smoothing, and over-squashing bottleneck multi-hop vulnerability predictions.

This research develops an Enhanced Dynamic GCN that introduces:

• Temporal Edge Update Protocol: Dynamically recomputes adjacency weights using learned node states and exponential memory decay.
• Post-Initialization Exogenous Shock Integration: Injects real-world rainfall variability and demand spike vectors directly into tier-specific node embeddings.
• Direction-Aware Gated Residual Propagation: Decomposes graph convolutions into forward supply and reverse demand flows with gated residual connections and multi-layer embedding preservation.

---

Run locally

The app runs on your machine: a FastAPI backend and a Next.js frontend. The helper script starts both.

powershell -File scripts\start-app.ps1

It launches the backend (uvicorn via the project .venv) on http://127.0.0.1:8000 and the frontend dev server on http://127.0.0.1:3000. The backend warms up its trained model suite on startup; wait for that to finish before running a simulation. See frontend/README.md for the frontend-only commands and the wizard walkthrough.