Simple Contrastive Graph Clustering.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Data Mining

Background:

• Contrastive learning shows promise in deep graph clustering.
• Existing methods suffer from inefficient data augmentations and graph convolutional operations.

Purpose of the Study:

• To propose a Simple Contrastive Graph Clustering (SCGC) algorithm.
• To improve efficiency and performance in deep graph clustering.

Main Methods:

• Introduced a simplified network architecture with preprocessing and a two-multilayer perceptron (MLP) backbone.
• Developed a novel data augmentation strategy by perturbing node embeddings directly using parameter-unshared Siamese encoders.
• Designed a cross-view structural consistency objective function to enhance discriminative capabilities.

Main Results:

• SCGC demonstrates effectiveness and superiority across seven benchmark datasets.
• Achieved at least a seven-fold average speedup compared to recent contrastive deep clustering competitors.
• Validated through extensive experimental results.

Conclusions:

• SCGC offers an efficient and effective approach to deep graph clustering.
• The proposed methods for architecture, data augmentation, and objective function significantly improve performance and speed.
• SCGC represents a notable advancement in the field of graph representation learning for clustering.