Africa-Neighbourhood-GCN

African Geo-GCN: Graph Neural Networks on Geopolitical Data

The Concept

This project explores the power of Graph Convolutional Networks (GCNs) applied to a custom-built dataset of African nations.

Instead of treating countries as isolated data points (standard Machine Learning), we model the continent as a Graph where:

• Nodes are countries.
• Edges represent shared land borders.
• Node Features are economic/demographic indicators (GDP, Population).

The goal is to perform Semi-Supervised Node Classification to predict a country's geopolitical region based on its neighbors' influence, demonstrating how GNNs aggregate spatial information.

Architecture (From Scratch)

This implementation avoids high-level GNN libraries to focus on the mathematical mechanics of the Kipf & Welling (2017) propagation rule:

$$ H^{(l+1)} = \sigma(\tilde{D}^{-\frac{1}{2}}\tilde{A}\tilde{D}^{-\frac{1}{2}} H^{(l)} W^{(l)}) $$

Project Structure

• src/models.py: Custom PyTorch implementation of the Graph Convolution Layer.
• src/data.py: Manual construction of the African Adjacency Matrix and Feature set.
• src/utils.py: Spectral normalization of the adjacency matrix.

How to Run

1. Install Dependencies

   pip install -r requirements.txt

2. Execute the Training

   python main.py

Results

The model is trained on a subset of countries and tested on "hidden" nodes (e.g., Kenya, Togo) to verify generalization.

Input Features: GDP ($B), Population (M)

Graph: 15 Nodes, Undirected.

Test Accuracy: ~100% (Converges rapidly due to strong homophily in geopolitical borders).

Visualization

The training script automatically generates a visualization of the graph, coloring nodes by their predicted region.

---

Created with passion for AI & Africa.

JeffreyYAJ