Literature Survey

[JoelMathewC]

Starting an issue to look into resources on GNN models, Temporal GNN models, GPU optimizations etc. This is necessary because there seems to be a greater understanding of the domain to decide on frontend requirements (incase a redesign is to be done), correctness of middleware/backend implementation and possible optimization opportunities.

Reading material covered in this issue are as follows

• PyG-T
• ChebShev
• Temporal ChebShev
• DCRNN
• HGL
• Neugraph (SAGA-NN)

[JoelMathewC]

Pytorch Geometric Temporal

Paper Link: https://arxiv.org/abs/2104.07788
Status: Completed reading entire paper

Ideas for the project

1. Try implementing DCRNN, because according to the paper in page 4, they state that they had to write custom code for diffusion, since it wasn't there in PyG. Took a very look at the implementation code here and it seems like quite a bit of logic, would be nice to see if a vertex centric-program can model this.
2. The documentation updates every time the code base changes, that might help for us as well.
3. Additionally in their future work they talk about building models for continuous time models (snapshots not at regular intervals) and temporal models which operate on curved spaces such as hyperbolic and spherical spaces.

[JoelMathewC]

HGL: Accelerating Heterogeneous GNN Training with Holistic Representation and Optimization

Paper Link: https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=10046125
Implementation Link: https://github.com/ytgui/HGL-proto/tree/0.1.0
Status: Completed reading entire paper

Ideas for project

1. Make every graph a heterogenous graph, homogenous graphs are heterogenous graphs with 1 relation.
2. Numeros optimization techniques for that approach, maybe some of them could be adopted
3. Putting a pause on this now and looking into other branches.

[JoelMathewC]

TGL: A General Framework for Temporal GNN Training on Billion-Scale Graphs

Paper Link: https://arxiv.org/pdf/2203.14883.pdf

[JoelMathewC]

Closing this out here.