byamasu-patrick/graph-neural-network-lab

This repository is a personal learning project where the author builds several types of graph neural networks from scratch, rather than using an existing library's ready-made version. A graph neural network is a kind of AI model built to learn from data made up of connected items, like a social network of people or a network of linked documents, instead of the plain rows and columns most AI models expect. The project starts with the basic building blocks needed to work with graphs, such as representing which nodes connect to which, and how information gets passed between connected nodes. From there it implements three well known graph neural network designs: the Graph Convolutional Network, which aggregates information from neighboring nodes, the Graph Attention Network, which learns to weigh some neighbors as more important than others, and GraphSAGE, which is built to scale to very large graphs by sampling neighbors rather than using all of them at once. The author credits the original research papers behind each of these designs and lists them as references. The README states that future additions may include newer architectures such as Graph Isomorphism Networks, versions that handle graphs changing over time, graphs with multiple types of nodes, and graph based transformer models, though these are not implemented yet. Everything is written in Python using PyTorch for the neural network parts, along with NumPy for numerical work, NetworkX for handling graph structures, and Matplotlib for visualizing results. The code is organized into Jupyter notebooks for each architecture plus a folder of reusable source code for layers, models, and utility functions. This project is aimed at people already comfortable with machine learning who want to understand how graph neural networks work internally, rather than a general audience or complete beginners.