FastGAE: Scalable graph autoencoders with stochastic subgraph decoding.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Network Analysis

Background:

• Graph autoencoders (AE) and variational autoencoders (VAE) are effective for node embedding.
• Scalability remains a significant challenge for applying AE and VAE to large-scale graphs.
• Existing methods struggle to efficiently process graphs with millions of nodes and edges.

Purpose of the Study:

• To introduce FastGAE, a novel framework designed to overcome the scalability limitations of graph AE and VAE.
• To enable the application of AE and VAE methods to large real-world graphs.
• To significantly accelerate the training process for graph-based autoencoder models.

Main Methods:

• Development of FastGAE, a general framework for scaling graph AE and VAE.
• Implementation of a stochastic subgraph decoding scheme for efficient training.
• Empirical evaluation on diverse, large-scale real-world graph datasets.

Main Results:

• FastGAE achieves significant speedups in training graph AE and VAE models.
• The proposed method preserves or enhances the performance of node embeddings.
• Demonstrated superior performance compared to existing scalability solutions for graph AE/VAE.

Conclusions:

• FastGAE provides an effective solution for training graph autoencoders on massive graphs.
• The stochastic subgraph decoding approach is key to achieving scalability and performance.
• FastGAE represents a significant advancement in applying deep learning to large network data.