Deep graph reconstruction for multi-view clustering.

Area of Science:

• Machine Learning
• Data Mining
• Computer Science

Background:

• Graph-based multi-view clustering methods leverage graph embeddings for improved performance.
• Existing shallow models struggle to capture complex nonlinear information within multi-view data.

Purpose of the Study:

• To propose a novel deep graph reconstruction (DGR) framework for enhanced multi-view clustering.
• To address the limitations of shallow models in learning nonlinear information from multi-view data.

Main Methods:

• The proposed DGR framework integrates three modules: Multi-graph Fusion Module (MFM) for consensus graph generation, Graph Embedding Network (GEN) for node representation learning, and Clustering Assignment Module (CAM) for direct cluster assignment.
• A novel, powerful loss function is incorporated into the DGR framework.

Main Results:

• Extensive experiments on seven real-world datasets demonstrate the superior clustering performance of DGR.
• The DGR framework exhibits significant efficiency advantages over state-of-the-art methods.

Conclusions:

• The proposed deep graph reconstruction (DGR) framework offers a powerful and efficient solution for multi-view clustering.
• DGR effectively learns nonlinear information, leading to improved clustering accuracy.