Google AI Releases TensorFlow GNN 1.0 (TF-GNN): A Production-Tested Library for Building GNNs at Scale

Graph Neural Networks (GNNs) are deep learning methods that operate on graphs and are used to perform inference on data described by graphs. Graphs have been used in mathematics and computer science for a long time and give solutions to complex problems by forming a network of nodes connected by edges in various irregular ways. Traditional ML algorithms allow only regular and uniform relations between input objects, struggle to handle complex relationships, and fail to understand objects and their connections which is crucial for many real-world data.

Google researchers added a new library in TensorFlow, called TensorFlow GNN 1.0 (TF-GNN) designed to build and train graph neural networks (GNNs) at scale within the TensorFlow ecosystem. This GNN library is capable of processing the structure and features of graphs, enabling predictions on individual nodes, entire graphs, or potential edges.

In TF-GNN, graphs are represented as GraphTensor, a collection of tensors under one class consisting of all the features of the graphs — nodes, properties of each node, edges, and weights or relations between nodes. The library supports heterogeneous graphs, accurately representing real-world scenarios where objects and their relationships come in distinct types. In the case of large datasets, the graph formed has a high number of nodes and complex connections. To train these networks efficiently, TF-GNN uses the subgraph sampling technique in which a small part of the graphs is trained with enough of the original data to compute the GNN result for the labeled node at its center and train the model.

The core GNN architecture is based on message-passing neural networks. In each round, nodes receive and process messages from their neighbors, iteratively refining their hidden states to reflect the aggregate information within their neighborhoods. TF-GNN supports training GNNs in both supervised and unsupervised manners. Supervised training minimizes a loss function based on labeled examples, while unsupervised training generates continuous representations (embeddings) of the graph structure for utilization in other ML systems.

TensorFlow GNN 1.0 addresses the need for a robust and scalable solution for building and training GNNs. Its key strengths lie in its ability to handle heterogeneous graphs, efficient subgraph sampling, flexible model building, and support for both supervised and unsupervised training. By seamlessly integrating with TensorFlow’s ecosystem, TF-GNN empowers researchers and developers to leverage the power of GNNs for various tasks involving complex network analysis and prediction.