tkipf/gcn

This repository contains the original code for Graph Convolutional Networks (GCN), a machine learning technique introduced in a 2017 research paper by Thomas Kipf and Max Welling. The code is written in Python using TensorFlow, an older version of the framework. The paper was accepted at a major machine learning conference called ICLR 2017. The core idea is about classifying nodes in a graph when you only have labels for some of those nodes, a setup called semi-supervised learning. A graph here means any data that has items and connections between them: social networks where people are nodes and friendships are edges, or academic citation networks where papers are nodes and citations are edges. The model learns to assign each node a category by looking at both the node's own properties and the properties of its neighbors in the graph. To use it with your own data, you provide three pieces: a table describing which nodes are connected to which (an adjacency matrix), a table of features for each node, and a table of labels for the nodes you already know. The repository ships with three academic citation datasets (Cora, Citeseer, and Pubmed) so you can run the demo immediately without gathering data yourself. The code offers three model variants. The main GCN model is the one from the paper. A second variant uses a mathematical technique called Chebyshev polynomials for a different approach to spreading information across the graph. A third is a simple baseline network that does not use the graph structure at all, useful for comparison. There is also support for classifying entire graphs rather than individual nodes, which requires a slightly different data setup. This repository is primarily a research artifact. It is the reference implementation accompanying a specific paper, not a general-purpose library. Researchers and students studying graph-based machine learning commonly use it as a starting point or a benchmark.