Graph Neural Networks (GNNs) are a subfield of machine learning which has in recent years seen an explosion of research and development. Applicable to many fields, including social network analysis and molecular chemistry, GNNs have large appeal. As a result of this, the field is highly fragmented, with many substantially differing solutions to the problem of ”graph learning”. This thesis proposal (and subsequent thesis) focus on the implementation details of GNNs, and specifically the development of a GNN library for the rust language.

Rusts unique approach to performance, reliability and usability present a compelling foun- dation for developing a GNN library. The ownership system for memory management opens up new optimizations that recent libraries such as burn have taken advantage of and are demon- strating can lead to many significant and innovative optimisations. Rust machine learning libraries also offer unparalleled portability, supporting the web, embedded, and non-nvidia gpus as first class platforms.

The motivation for this work is twofold. Firstly, to fill the gap in the Rust ecosystem by providing a robust and efficient GNN library that leverages Rust’s unique features with a focus on convenience of API and portability. Secondly, to take advantage of the unique optimisations opened up by rust to enhance graph learning performance.