Self-Supervised Graph Representation Learning

This repository contains the implementation and experiments for the project Self-Supervised Graph Representation Learning, comparing supervised, self-supervised, and multi-task learning approaches on heterogeneous graph data.

Project Overview

We evaluate 3 architectures:

• Supervised GraphSAGE
• Self-supervised GraphMAE
• Multi-Task Learning (MTL) combining node and edge prediction from supervised

The models are trained and tested on the OGB-MAG dataset and evaluated on:

• Node classification (venue prediction)
• Link prediction (paper-field edges)

Key Findings

• Self-supervised GraphMAE produces embeddings competitive with supervised GraphSAGE.
• Multi-task learning improves node classification but slightly reduces link prediction performance.
• High AUC scores are influenced by strong class imbalance (large number of negative edges).

Repository Structure

/lib/                       # Our own python package where we store essentail functions and classes
/mag/                       # Processed OGB-MAG data and metapath2vec embeddings
/supervised/                # All train, test, models, and figures for supervised models
/supervised/models/         # Best model for supervised and MTL models
/batch/                     # Outputs and metrics from HPC

Report

The full project report is included as Self_supervised_Graph_Representation_Learning.pdf.