UNAGI: Unified neighbor-aware graph neural network for multi-view clustering

Area of Science:

• Machine Learning
• Data Mining
• Graph Neural Networks

Background:

• Multi-view graph refining-based clustering (MGRC) methods use Graph Neural Networks (GNNs) to learn data topology for clustering.
• Current MGRC methods employ a disjoint two-stage process, neglecting cross-view consistency and semantic information.

Purpose of the Study:

• To propose a Unified Neighbor-Aware Graph neural network for multi-view clustering (UNAGI).
• To address limitations of existing MGRC methods by integrating graph topology optimization and sample representation learning.

Main Methods:

• Developed a novel framework merging graph topology optimization and sample representations via a differentiable graph adapter for unified training.
• Introduced a regularization technique for robust graph learning and inter-view graph topology alignment using neighbor-aware pseudo-labels.

Main Results:

• UNAGI demonstrated superior clustering performance across seven diverse datasets.
• The unified training paradigm and novel regularization significantly enhanced clustering accuracy.

Conclusions:

• UNAGI offers an effective solution for multi-view clustering by overcoming limitations of previous MGRC approaches.
• The proposed method showcases the benefits of integrated graph learning and representation learning for complex data.