Parallelly Adaptive Graph Convolutional Clustering Model.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Mining

Background:

• Graph convolutional networks (GCNs) excel at tasks leveraging data topology.
• GCN performance is limited by noisy or outlier-corrupted graph structures.
• Existing methods often rely on fixed, pre-trained graphs.

Purpose of the Study:

• To propose a novel end-to-end adaptive graph convolutional clustering (AGCC) model.
• To address the limitations of fixed graphs in GCN-based clustering.
• To enhance clustering accuracy by learning adaptive graph structures from data.

Main Methods:

• Developed a parallel two-pathway network: an adaptive graph convolutional (AGC) module and an auto-encoder (AE) module.
• The AGC module iteratively updates graph structure and data representation.
• An attention-mechanism-based fusion (AMF) module integrates representations from AGC and AE, mitigating over-smoothing.

Main Results:

• The proposed AGCC model demonstrated superior performance across six public datasets.
• Experimental results validate the effectiveness of the adaptive graph learning approach.
• AGCC outperformed several state-of-the-art deep clustering methods.

Conclusions:

• The AGCC model effectively overcomes the limitations of fixed graphs in GCNs for clustering.
• Adaptive graph learning significantly improves clustering performance.
• The novel architecture provides a robust solution for deep clustering tasks.