One toolkit, many node types: a heterogeneous GNN workbench

What it does OpenHGNN is a Python toolkit built on DGL and PyTorch that collects a wide range of heterogeneous graph neural network models under one roof. It wraps published architectures—RGCN, HGT, HAN, GTN, and many others—into a unified interface for node classification, link prediction, and recommendation tasks. The project also folds in hyperparameter optimization via Optuna, mini-batch and distributed training support, and a growing benchmark suite.

The interesting bit Instead of reimplementing attention mechanisms for multi-relational graphs from scratch, it treats the academic literature as a modular parts bin. You can swap models, tasks, or datasets, or inject your own into the pipeline without forking the entire framework. That is rarer than it sounds in graph ML tooling.

Key highlights

• Ships with implementations of a large catalog of published models (RGCN, HAN, HGT, GTN, SimpleHGN, etc.) across node classification, link prediction, and recommendation.
• Recent releases added graph prompt pipelines, dgl.graphBolt data processing, sparse aggregators, and distributed training.
• Integrates Optuna for hyperparameter search out of the box.
• Supports custom datasets and models, plus optional graph database backends (Neo4j, NebulaGraph) via gdbi.
• Includes a benchmark leaderboard and TensorBoard visualization.

Verdict Researchers and practitioners working with multi-relational or multi-typed graph data who want to compare published methods or prototype variants should look here. If your graphs are homogeneous and simple, the full heterogeneous machinery is likely overkill.