Improved Normalized Cut for Multi-View Clustering.

Area of Science:

• Computer Science
• Data Science
• Machine Learning

Background:

• Spectral clustering (SC) excels at grouping complex data structures.
• Traditional SC methods use sequential spectral decomposition and rounding, which can lead to suboptimal results.
• The rounding stage in conventional SC may cause significant deviation from true cluster assignments.

Purpose of the Study:

• To develop a unified framework for simultaneously learning optimal continuous and binary indicator matrices in clustering.
• To address limitations of traditional two-stage spectral clustering approaches.
• To enhance multi-view clustering while retaining capability for single-view scenarios.

Main Methods:

• A general joint framework is proposed to learn optimal continuous and binary indicator matrices concurrently.
• Theoretical proofs are provided for the developed method.
• An effective alternating updating algorithm is designed for optimizing the objective function.

Main Results:

• The proposed method demonstrates superior performance across six clustering metrics on benchmark datasets.
• Empirical results show outperformance compared to several state-of-the-art methods.
• The joint optimization framework effectively mitigates issues associated with traditional two-stage SC.

Conclusions:

• The novel joint framework offers a more effective approach to spectral clustering, particularly for multi-view data.
• Simultaneous learning of continuous and binary matrices leads to improved clustering accuracy.
• The method provides theoretical guarantees and practical advantages over existing techniques.