Deep graph clustering via aligning representation learning.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Data Mining

Background:

• Deep graph clustering is crucial for analyzing graph-structured data.
• Existing autoencoder and graph convolutional network methods often yield non-discriminative node representations.
• Current contrastive graph clustering methods are limited by reliance on data augmentation and lack of self-consistency.

Purpose of the Study:

• To propose a novel contrastive deep graph clustering method, Aligning Representation Learning Network (ARLN).
• To enhance node representation discriminability and clustering performance.
• To address limitations of existing methods regarding data augmentation and self-consistency.

Main Methods:

• Utilizing contrastive learning between an autoencoder and a graph autoencoder to bypass complex data augmentations.
• Introducing instance and feature contrastive modules for consensus representation learning.
• Designing an assignment probability contrastive module to ensure self-consistency between node representations and cluster assignments.

Main Results:

• ARLN learns discriminative node representations through contrastive learning.
• The method maintains self-consistency between node representations and cluster assignments.
• Experimental results demonstrate the superiority of ARLN over state-of-the-art deep graph clustering methods on benchmark datasets.

Conclusions:

• ARLN offers an effective approach to deep graph clustering by leveraging contrastive learning.
• The proposed method improves representation learning and clustering accuracy.
• ARLN provides a robust alternative to existing methods, particularly those dependent on data augmentation.