Graph Neural Networks Projects

This repository contains implementations of state-of-the art graph machine learning models for recommendations, link prediction and node classification:

1. Movie recommendations with IGMC [1] (link prediction). Movies are recommended to users using only information from subgraphs.
2. Link prediction using SEAL [2]. Predict missing links in a citation network by using the SEAL framework.
3. Node classification using HGT [3]. Predict node properties in a heterogeneous graph.

Installation requirements

ogb>=1.3.0
torch>=1.11.0
pytorch-lightning>=1.2.0
torch-geometric==master (pip install git+https://github.com/rusty1s/pytorch_geometric.git)

Other required python libraries include: numpy, scipy, tqdm etc. These are found in requirements.txt.

Performance

Recommendations (link prediction)

Model	Dataset	Valid RMSE	Test RMSE	#Parameters
IGMC	MovieLens	52.67	52.73	49k

Node classification

Model	Dataset	Valid Accuracy (%)	Test Accuracy (%)	#Parameters
HGT	Cora	92.3	-	19M

Link prediction

Model	Dataset	Valid AP / AUROC	Test AP / AUROC	#Parameters
HGT	Cora	92.3	-	19M
SEAL	Cora	94.3	92.6	96K

References

[1] M. Zhang and Y. Chen: Inductive Matrix Completion Based On Graph Neural Networks

[2] M. Zhang and Y. Chen: Link Prediction Based on Graph Neural Networks

[3] Z. Hu et al.: Heterogeneous Graph Transformer