Multiscale Contrastive Learning for Node Clustering Based on Variational Graph Auto-Encoder.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Unsupervised Learning

Background:

• Variational Graph Auto-Encoders (VGAEs) are widely used for node clustering.
• Existing VGAEs suffer from issues like posterior collapse (PC), feature randomness (FR), and feature drift (FD).
• These challenges stem from mismatches in inference/generative models and noisy clustering assignments.

Purpose of the Study:

• To propose a novel Multiscale Contrastive Variational Graph Auto-Encoder (MCVGAE) to overcome limitations in current VGAE node clustering.
• To improve the alignment between latent representations and data distribution, preventing PC.
• To effectively reduce FR and FD for enhanced clustering performance.

Main Methods:

• MCVGAE integrates cluster-level and graph-level contrastive learning.
• It employs proximity-level and cluster-level self-supervised learning strategies.
• The multiscale approach enhances representation learning and clustering accuracy.

Main Results:

• MCVGAE demonstrated superior performance across multiple benchmark datasets (Cora, ACM, Pubmed, Citeseer, DBLP, Wiki).
• Achieved high accuracy scores, e.g., 79.09% on Cora and 90.04% on ACM.
• Outperformed 30 state-of-the-art node clustering methods.

Conclusions:

• MCVGAE effectively addresses critical challenges in VGAE-based node clustering.
• The proposed model offers improved latent space alignment and robust feature representation.
• MCVGAE represents a significant advancement in graph representation learning for clustering tasks.