How tkipf/gae works
Project type and purpose
• This is a research-oriented machine learning library implementing Graph Auto-Encoders (GAEs) and Variational Graph Auto-Encoders (VGAEs) in TensorFlow.
• Its primary function is to load graph data (cora, citeseer, pubmed), build a GCN-based encoder and decoder, optimize a reconstruction or variational loss, and produce embeddings or link-prediction scores.
File structure
Top-level files
• setup.py – packaging and install script
• README.md, LICENCE, .gitignore
g a e/
• data/ – serialized dataset files (adjacency + feature matrices for cora, citeseer, pubmed)
• input_data.py – data-loading utilities
• preprocessing.py – adjacency normalization and feature preprocessing
• initializations.py – weight and bias initializers
• layers.py – Graph Convolutional Network (GCN) layer definitions
• model.py – encoder/decoder graph construction for AE and VAE
• optimizer.py – loss functions (reconstruction, KL) and optimizer setup
• train.py – end-to-end training script, hyperparameter parsing
README for additional insights
• Describes two models: gcn_ae and gcn_vae
• Specifies required inputs: N×N adjacency, N×D feature matrices
• Recommends how to swap datasets or models via command-line flags
Elements
Data files → input_data → preprocessing → model → optimizer → outputs/logs
“Adjacency + Feature matrices”
“Normalized adjacency”
“Latent embeddings z”
“Reconstruction loss / KL divergence”