Spektral is an open-source library for graph neural networks built on Keras and TensorFlow 2. Released under the MIT license with nearly 2,400 GitHub stars, its last updates landed in 2024, and it remains a widely used entry point for researchers who want to experiment with deep learning on graphs without leaving the Keras workflow.
Graph neural networks power problems where data is relational โ molecules, social networks, citation graphs, recommendation. Spektral provides the building blocks: dozens of convolutional and pooling layers (GCN, GraphSAGE, GAT, ChebNet and more), loaders for benchmark datasets, and utilities for the graph-specific data handling that ordinary deep learning frameworks ignore. Because it is built on Keras, models compose with the rest of the TensorFlow ecosystem, and the library’s design goal is that a graph model should read like any other Keras model.
Spektral is free and open source under the MIT license, with no commercial edition.
Spektral suits researchers, students and data scientists exploring graph machine learning โ molecular property prediction, node classification, link prediction โ who prefer Keras ergonomics over lower-level graph frameworks.
The library’s website and documentation include worked examples that span the main graph learning problem types: node classification on citation networks such as Cora, graph classification on molecular and protein benchmarks, link prediction, and graph generation. Spektral’s layers are implemented to follow the original papers closely, which has made it a popular teaching and prototyping tool in graph machine learning courses as well as a baseline implementation that newer research compares against.