Graph Neural Networks (GNN) are deep learning models well adapted to data that takes the form of graphs with feature vectors associated to nodes and edges. GNNs are a growing area of research and find many applications in complex networks analysis, relational reasoning, combinatorial optimization, molecule generation, and many other fields.
GraphNeuralNetworks.jl is a pure Julia package for GNNs equipped with many features. It implements common graph convolutional layers, with CUDA support and graph batching for fast parallel operations. There are a number of ways by which the package could be improved.
While we implement a good variety of graph convolutional layers, there is still a vast zoology to be implemented yet. Preprocessing tools, pooling operators, and other GNN-related functionalities can be considered as well.
Duration: 175h.
Expected difficulty: easy to medium.
Expected outcome: Enrich the package with a variety of new layers and operators.
As part of the documentation and for bootstrapping new projects, we want to add fully worked out examples and applications of graph neural networks. We can start with entry-level tutorials and progressively introduce the reader to more advanced features.
Duration: 175h.
Expected difficulty: medium.
Expected outcome: A few pedagogical and more advanced examples of graph neural networks applications.
Provide Julia friendly wrappers for common graph datasets in MLDatasets.jl. Create convenient interfaces for the Julia ML and data ecosystem.
Duration: 175h.
Expected difficulty: easy.
Expected outcome: A large collection of graph datasets easily available to the Julia ecosystem.
In some complex networks, the relations expressed by edges can be of different types. We currently support this with the GNNHeteroGraph type but none of the current graph convolutional layers support heterogeneous graphs as inputs. With this project we will implement a few layers for heterographs.
Duration: 350h.
Expected difficulty: hard.
Expected outcome: The implementation of a new graph type for heterogeneous networks and corresponding graph convolutional layers.
Graph containing several millions of nodes are too large for gpu memory. Mini-batch training is performed on subgraphs, as in the GraphSAGE algorithm.
Duration: 175h.
Expected difficulty: hard.
Expected outcome: The necessary algorithmic components to scale GNN training to very large graphs.
We aim at implementing temporal graph convolutions for time-varying graph and/or node features. The design of an efficient dynamical graph type is a crucial part of this project.
Duration: 350h.
Expected difficulty: hard.
Expected outcome: A new dynamical graph type and corresponding convolutional layers.
Many graph convolutional layers can be expressed as non-materializing algebraic operations involving the adjacency matrix instead of the slower and more memory consuming gather/scatter mechanism. We aim at extending as far as possible and in a gpu-friendly way these fused implementation.
Duration: 175h.
Expected difficulty: hard.
Expected outcome: A noticeable performance increase for many graph convolutional operations.
Familiarity with graph neural networks and Flux.jl.
Carlo Lucibello (author of GraphNeuralNetworks.jl). For linear algebra, co-mentoring by Will Kimmerer (lead developer of SuiteSparseGraphBLAS.jl). Feel free to contact us on the Julia Slack Workspace or by opening an issue in the GitHub repo.