pyg-team/pytorch_geometric

PyTorch Geometric, often called PyG, is a Python library for building and training Graph Neural Networks. To understand why that matters, think of data that is naturally structured as a network: a social graph of friends, a citation network of research papers, or the atoms in a molecule connected by bonds. Standard deep learning tools like image classifiers work on grids of pixels, but they cannot easily process this kind of interconnected data. Graph Neural Networks are a family of AI techniques designed specifically for graphs, and PyG makes them straightforward to use. The library is built on top of PyTorch, so anyone already familiar with PyTorch will find the interface familiar. It provides dozens of pre-built graph network layer types taken directly from published research papers, plus tools for loading standard benchmark datasets, splitting data into batches, and applying transformations to graph data before training. You can also build custom network layers by defining how information passes between connected nodes in the graph, a design pattern the library calls message passing. PyG handles graphs of many shapes: small graphs in large batches (like molecules in drug discovery), single enormous graphs with millions of nodes (like social networks), and dynamic graphs that change over time. You would use it if you are a machine learning researcher or engineer working on problems where relationships between entities matter, such as recommendation systems, molecular property prediction, fraud detection in transaction networks, or knowledge graph reasoning. The tech stack is Python and PyTorch.