Structure-aware deep clustering network based on contrastive learning.

Area of Science:

• Artificial Intelligence
• Data Mining
• Machine Learning

Background:

• Deep clustering methods, including auto-encoder (AE) and graph neural network (GNN) based approaches, are widely used.
• Existing AE methods struggle with structural information extraction, while GNNs face issues like smoothing and heterophily.
• Combined AE-GNN methods show promise but lack a balance between preserving raw and exploring underlying data structures.

Purpose of the Study:

• To propose a novel Structure-Aware Deep Clustering network (SADC).
• To enhance the extraction of both raw and underlying data structures in deep clustering.
• To improve the performance of deep clustering tasks by addressing limitations of existing methods.

Main Methods:

• Computing cumulative influence of non-adjacent nodes to enhance the adjacency matrix.
• Designing an enhanced graph auto-encoder.
• Endowing the AE latent space with raw structure perception capabilities.
• Implementing self-supervised mechanisms for co-optimizing node representation and topology learning.
• Developing a novel loss function for preserving inherent structure while exploring latent data structure.

Main Results:

• The proposed SADC network effectively balances preserving raw data structure and exploring latent structures.
• Experiments on six benchmark datasets demonstrate superior performance compared to state-of-the-art methods.
• The enhanced adjacency matrix and graph auto-encoder contribute to improved clustering accuracy.

Conclusions:

• SADC offers a significant advancement in deep clustering by effectively integrating structural information.
• The method provides a balanced approach to structure preservation and exploration, leading to better data mining outcomes.
• SADC represents a promising direction for future research in deep clustering and graph-based learning.