DIG: A Turnkey Library for Diving into Graph Deep Learning Research (2103.12608v3)

Abstract: Although there exist several libraries for deep learning on graphs, they are aiming at implementing basic operations for graph deep learning. In the research community, implementing and benchmarking various advanced tasks are still painful and time-consuming with existing libraries. To facilitate graph deep learning research, we introduce DIG: Dive into Graphs, a turnkey library that provides a unified testbed for higher level, research-oriented graph deep learning tasks. Currently, we consider graph generation, self-supervised learning on graphs, explainability of graph neural networks, and deep learning on 3D graphs. For each direction, we provide unified implementations of data interfaces, common algorithms, and evaluation metrics. Altogether, DIG is an extensible, open-source, and turnkey library for researchers to develop new methods and effortlessly compare with common baselines using widely used datasets and evaluation metrics. Source code is available at https://github.com/divelab/DIG.

• The paper introduces DIG, a unified and extensible Python library designed to streamline advanced research in graph deep learning.
• The paper implements cutting-edge algorithms for graph generation, self-supervised learning, and explainability, benchmarked on datasets like QM9 and ZINC250k.
• The paper offers a modular design that enhances reproducibility and customization, lowering the entry barrier for complex GNN research applications.

Overview of DIG: A Turnkey Library for Graph Deep Learning

This paper presents DIG (Dive into Graphs), a comprehensive Python library designed to support advanced research and experimentation in graph deep learning. While various libraries already exist to facilitate basic operations and tasks in graph neural networks (GNNs), DIG addresses the need for a more unified and extensible framework targeted at higher-level, research-oriented tasks. This library provides researchers with the resources to implement, benchmark, and compare advanced algorithms across multiple directions such as graph generation, self-supervised learning, explainability, and 3D graph analysis.

Key Contributions and Features

DIG stands out for its focus on advancing research methodologies, offering a suite of implementations for tasks that go beyond basic node and graph classification. The library's primary contributions can be summarized as follows:

1. Graph Generation: DIG implements advanced graph generation algorithms, crucial for applications like molecule discovery. Supported algorithms include JT-VAE, GraphAF, GraphDF, and GraphEBM, with interfaces for datasets such as QM9 and ZINC250k.
2. Self-Supervised Learning: The library broadens the scope to self-supervised learning on graphs, incorporating algorithms such as InfoGraph, GRACE, MVGRL, and GraphCL. The library provides interfaces for both node-level and graph-level tasks, facilitating expressive representation learning.
3. Explainability: Recognizing the importance of understanding model predictions, DIG includes tools for explainability in GNNs, including GNNExplainer, PGExplainer, DeepLIFT, and others. This capability is essential for deploying models in real-world applications where transparency is critical.
4. Deep Learning on 3D Graphs: DIG supports algorithms tailored for 3D graphs, leveraging the 3DGN framework to unify implementations like SchNet and DimeNet++. This aspect is significant for domains where spatial information is paramount, such as molecular chemistry.

Design Considerations

DIG is engineered with a focus on unified implementation, extensibility, and customization:

• Unified Implementation: By providing standardized APIs for data interfaces, algorithms, and evaluation metrics, DIG ensures consistency across tasks, simplifying the integration of diverse algorithms within a uniform framework.
• Extensibility and Customization: The modular design allows easy incorporation of new datasets and algorithms, enabling researchers to customize and extend their experimental setups effortlessly.

Quality and Documentation

DIG adheres to high-quality standards essential for open-source software. The library ensures code reliability and reproducibility, with continuous integration testing and extensive documentation, which includes detailed API descriptions and practical tutorials. Additionally, it encourages community contributions under a permissive GNU GPLv3 license.

Conclusion

DIG emerges as a pivotal library for researchers aiming to explore advanced facets of graph deep learning. By offering a turnkey solution with unified, extensible, and customizable components, it significantly reduces the entry barrier for experimenting with cutting-edge methodologies in the domain. Future developments may see further expansions and algorithm additions, continually aligning with the evolving landscape of graph deep learning research.