Structure enhanced deep clustering network via a weighted neighbourhood auto-encoder

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Deep clustering integrates semantic and structural data using neural networks.
• Existing methods using auto-encoders (AE) and graph convolutional networks (GCNs) face issues with structural information loss due to GCN over-smoothing.

Purpose of the Study:

• To propose a structure-enhanced deep clustering network to address the vanishing structural information problem.
• To improve feature representation and clustering performance by enhancing structural data within the learning process.

Main Methods:

• A novel structure-enhanced AE, the weighted neighborhood AE (wNAE), is introduced to learn structure-enhanced semantic (SES) representations.
• The GCN-specific structural representation is enhanced and supervised by its own structural information.
• A joint supervision strategy is designed for simultaneous learning of wNAE, GCN modules, and clustering assignments.

Main Results:

• The proposed network effectively enhances GCN-specific structural data representation by integrating SES representations.
• Experimental results demonstrate the importance of preserving both semantic and neighbor-wise structural information for effective clustering.
• The method empirically validates improved clustering performance across various datasets.

Conclusions:

• Preserving structural information is crucial for deep clustering tasks, particularly when using GCNs.
• The proposed structure-enhanced deep clustering network, incorporating wNAE, offers a robust solution to mitigate information loss.
• This approach advances deep clustering by ensuring more representative features and superior clustering accuracy.