Posted On: Feb 11, 2022

You can now define custom machine learning models with Neptune Machine Learning (ML) for your graph data on Amazon Neptune. Neptune ML is Amazon Neptune’s machine learning capability that uses Graph Neural Networks (GNNs) developed with Deep Graph Library (DGL) to automate the heavy lifting of selecting and training ML models for graph data. With this launch, you can also run SPARQL inference queries on W3C’s Resource Description Framework (RDF) data model, in addition to Apache TinkerPop Gremlin inference queries on property graphs. New machine learning tasks for RDF include object classification, object regression, object prediction and subject prediction.

Custom model training is intended for users who want to bring their own custom GNN models developed in DGL, or for advanced use cases in node classification and regression, such as using tabular or ensemble models. For example, you can build a custom ML model to link customer records in an identity graph or combine the predictions from a non-graph and a graph model for fraud detection. With SPARQL support, Neptune ML can infer categorical classification or numerical regression on the properties of both objects and subjects. Neptune ML can also predict the most likely object given an existing subject and predicate, and vice versa on RDF data.

You can use the quick-start setup to get started with Neptune ML. Neptune ML is available from Neptune versions 1.0.5.0 and later, in all regions where Amazon Neptune is available. There are no additional charges for using Neptune ML. You only pay for the resources provisioned, such as Amazon Neptune, Amazon SageMaker, Amazon CloudWatch, and Amazon S3.

For more information on custom models, see the documentation or check out the example models for node classification and other tasks on GitHub. Sample SPARQL inference queries for Neptune ML are available in our documentation.