Awesome

Revisiting and Benchmarking Graph Autoencoders: A Contrastive Learning Perspective

We introduce the lrGAE (left-rigt GAE) benchmark --- graph autoencoders as contrastive learning architectures. lrGAE provides a new contrastive learning perspective of designing powerful GAEs from five dimensions:

• Augmentations
• Encoder/decoder networks
• Contrastive views
• Contrastive losses
• (optional) Negative examples

The contrastive views is the key to design different yet advanced GAEs, which invovels three components: graph views, receptive fields, and node pairs.

• Graph views: the graph or augmentated graph in two contrastive views, denoted as graph $A$ or $B$.
• Receptive fields: the depth of the graph neural network or the number of sampled hops in a node's neighborhood, denoted as $l$ or $r$.
• Node pairs: the contrasting objective over a single node $v$ or two nodes $v$ and $u$.

Therefore, we have $2^3=8$ variants of lrGAE in terms of the contrastive views, shown below:

[!NOTE] Actually, we got 7 variants since lrGAE-1 is not applicable as a contrastive method. There are more than 7 variants of lrGAE, you can design more powerful GAEs by exploring different combinations of augmentation strategies, encoder/decoder networks, contrastive views, contrastive losses and even the negative sampling tricks.

💫 Environment Setup

Before you begin, please make sure that you have Anaconda or Miniconda installed on your system. This guide assumes that you have a CUDA-enabled GPU.

# Create and activate a new Conda environment named 'lrGAE'
conda create -n lrGAE python==3.12 -c conda-forge -y
conda activate lrGAE

# Install Pytorch 2.3.1 with CUDA 12.1 support
# If your use a different CUDA version, please refer to the PyTorch website for the appropriate versions.
pip install torch==2.3.1 torchvision==0.18.1 torchaudio==2.3.1 --index-url https://download.pytorch.org/whl/cu121

# Install PyG
pip install torch_geometric
# Install additional dependencies of PyG
pip install pyg_lib torch_scatter torch_sparse torch_cluster -f https://data.pyg.org/whl/torch-2.3.0+cu121.html

Additional dependences of PyG can be found at https://pytorch-geometric.readthedocs.io/en/latest/notes/installation.html

🚀 Installation

Please make sure you have installed PyTorch and PyTorch Geometric (PyG).

# Coming soon
pip install -U lrgae

or

# Recommended
git clone https://github.com/EdisonLeeeee/lrGAE.git && cd lrGAE
pip install -e . --verbose

where -e means "editable" mode so you don't have to reinstall every time you make changes.

📍 Reproduction

Five graph-based learning tasks are supported:

• Node classification task
• Link prediction task
• Graph clustering task
• Graph classification task
• Heterogeneous node classification task

👀 Implementations

• GAE: Variational graph auto-encoders. NeurIPS 2016
• MaskGAE: What’s behind the mask: Understanding masked graph modeling for graph autoencoders. KDD 2023
• GraphMAE: GraphMAE: Self-supervised masked graph autoencoders. KDD 2022
• GraphMAE2: GraphMAE2: A decoding-enhanced masked self-supervised graph learner. WWW 2023
• AUG-MAE: Rethinking graph masked autoencoders through alignment and uniformity. AAAI 2024
• GiGaMAE: GiGaMAE: Generalizable graph masked autoencoder via collaborative latent space reconstruction. CIKM 2023
• S2GAE: S2GAE: self-supervised graph autoencoders are generalizable learners with graph masking. WSDM 2023
• 7 variants of lrGAE in terms of different contrastive views (See Fig.2 for illustration)
  • 2️⃣ lrGAE-ABllvv
  • 3️⃣ lrGAE-AAlrvv
  • 4️⃣ lrGAE-ABlrvv
  • 5️⃣ lrGAE-AAllvu
  • 5️⃣ lrGAE-AAlrvu
  • 7️⃣ lrGAE-ABllvu
  • 8️⃣ lrGAE-ABlrvu