Source-linked AI summary
DIG: A Turnkey Library for Diving into Graph Deep Learning Research
Meng Liu, Youzhi Luo, Limei Wang, Yaochen Xie, Hao Yuan, Shurui Gui, Haiyang Yu, Zhao Xu, Jingtun Zhang, Yi Liu, Keqiang Yan, Haoran Liu, Cong Fu, Bora Oztekin, Xuan Zhang, Shuiwang Ji
TL;DR
Existing graph libraries emphasize basic operations, while implementing and benchmarking advanced graph-learning tasks remains costly. DIG offers unified components across four research directions, and its benchmark examples reproduce original-paper results within reasonable or negligible differences.
Problem
Existing graph libraries focus largely on basic components and elementary tasks, while implementing and benchmarking advanced graph-learning tasks remains effortful.
Source: §1 · p. 2Method
DIG provides unified, extensible data interfaces, common algorithms, and evaluation metrics for graph generation, graph self-supervised learning, GNN explainability, and deep learning on 3D graphs.
Source: §1 · p. 2Results
Benchmark examples for advanced algorithms reproduce the original papers’ reported results within reasonable or negligible differences.
Source: §4 · p. 4Takeaways & Limitations
DIG can serve as a platform for developing and benchmarking algorithms in its covered directions, with datasets and metrics that users can customize.
Source: §3 · p. 4
Abstract
from arXiv · showhide
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.
1. Introduction
Existing graph-learning libraries focus mainly on basic components and elementary tasks, leaving advanced-task implementation and benchmarking costly. DIG provides unified, extensible tools across four research directions.
1. Introduction
- Existing libraries focus mainly on basic graph-neural-network components and elementary tasks, while implementing and benchmarking advanced tasks remains effortful.The passage gives graph generation as an example of an advanced task.Source: §1 · p. 2
- DIG covers graph generation, graph self-supervised learning, GNN explainability, and deep learning on 3D graphs with unified, extensible data interfaces, algorithms, and metrics.The library is presented as a turnkey, open-source testbed for developing methods and comparing them with common baselines on widely used datasets and metrics.Source: §1 · p. 2
2. Library Description
DIG packages algorithms, data interfaces, and evaluation metrics for four graph-learning research directions. Its library description specifies representative methods, datasets, and task-specific evaluation APIs.
2. Library Description
- DIG is built with Python and PyTorch, incorporates some PyG and RDKit operations, and currently includes four directions and 18 algorithms.The library is designed to accommodate additional directions and algorithms.Source: §2 · p. 2
- For molecular graph generation, DIG includes JT-VAE, GraphAF, GraphDF, and GraphEBM, with QM9, ZINC250k, and MOSES interfaces and APIs for three generation-evaluation settings.Those settings are random generation, property optimization, and constrained property optimization.Source: §2 · p. 2
- Graph self-supervised learning includes InfoGraph, GRACE, MVGRL, and GraphCL, with graph-level and node-level classification data interfaces and standard classification metrics.The interfaces cover TUDataset and citation networks including Cora, CiteSeer, and PubMed.Source: §2 · pp. 2–3
- GNN explainability includes seven explanation algorithms, synthetic and molecule datasets, text-derived human-understandable graph datasets, and Fidelity and Sparsity metrics.The named text-derived datasets include Graph-SST2 and Graph-SST5.Source: §2 · p. 3
- For 3D graphs, DIG includes SchNet, DimeNet++, and SphereNet in the 3DGN framework, with QM9 and MD17 interfaces and MAE evaluation.The passage defines 3D graphs as graphs whose nodes have associated 3D positions.Source: §2 · pp. 3–4
3. Key Design Considerations
DIG standardizes data interfaces, algorithms, and metrics while unifying implementations where methods share structure. Its extensibility and customization support incorporating components and configuring experiments.
3. Key Design Considerations
- DIG provides data-interface, algorithm, and metric APIs for each direction, creating a standardized testbed and unifying implementations where algorithms share non-trivial commonalities.The three 3D-graph algorithms share the 3DGN framework with different internal functions; many graph self-supervised methods are treated as contrastive models with unified objective functions.Source: §3 · p. 4
- Users can add datasets, algorithms, and metrics, then choose data interfaces and evaluation metrics for customized experiments.The paper describes DIG as a platform for implementing and benchmarking algorithms in its covered directions.Source: §3 · p. 4
4. Quality Standards
DIG assesses software quality through reliability and reproducibility, documentation, and openness. Its tested APIs and benchmark examples support reproducing original-paper results within reasonable or negligible differences.
4. Quality Standards
- Travis CI extensively tests data-interface and evaluation-metric APIs, while benchmark examples reproduce advanced-algorithm results from original papers within reasonable or negligible differences.Source: §4 · p. 4
- Online documentation provides detailed API descriptions and hands-on tutorials.Source: §4 · p. 4
- DIG welcomes community contributions under documented contribution guidelines and is distributed under the GNU GPLv3 license.Source: §4 · p. 4
5. Conclusion and Outlook
DIG unifies extensible implementations of data interfaces, common algorithms, and evaluation metrics across four graph deep learning research directions. The authors hope it will ease algorithm development and benchmarking, and plan to add emerging directions and algorithms.
- DIG provides unified, extensible implementations of data interfaces, common algorithms, and evaluation metrics across graph generation, self-supervised learning, GNN explainability, and deep learning on 3D graphs.Source: §5 · pp. 4–5
- The authors hope DIG will enable researchers to implement and benchmark algorithms more easily.Source: §5 · p. 5
- The authors plan to incorporate more emerging research directions and advanced algorithms into DIG.Source: §5 · p. 5