Adaptive graph convolutional clustering network with optimal probabilistic graph.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Mining

Background:

• Graph convolutional network (GCN)-based clustering excels at leveraging topological structure.
• Existing methods often construct fixed graphs independently, leading to suboptimal embeddings with noisy data.
• The reliability of the adjacency graph significantly impacts clustering performance, especially for non-graph data.

Purpose of the Study:

• To propose an adaptive graph convolutional clustering network (AGCCN) for improved deep graph clustering.
• To develop a unified framework that simultaneously learns graph structure and node embedding representations.
• To enhance the robustness and effectiveness of GCN-based clustering methods.

Main Methods:

• Developed an adaptive graph convolutional clustering network (AGCCN) that unifies graph structure learning and node embedding.
• Learned a weighted adjacency graph adaptively from node representations via an optimization problem, assigning probabilistic neighbors.
• Fused attribute features from a parallel Auto-Encoder (AE) module and employed a dual self-supervised clustering mechanism.

Main Results:

• The proposed AGCCN adaptively constructs reliable similarity graphs, overcoming limitations of fixed graph structures.
• The method effectively learns comprehensive node embeddings by integrating attribute features and adaptive graph convolutions.
• Experimental results on real-world datasets demonstrate the superiority and effectiveness of the AGCCN method over existing approaches.

Conclusions:

• The AGCCN provides a robust and effective deep graph clustering method by adaptively learning graph structures.
• The unified framework alleviates the over-smoothing problem in GCNs and enhances representation discriminability.
• The proposed approach significantly improves clustering performance across various real-world datasets.