Discrete and Balanced Spectral Clustering With Scalability.

Area of Science:

• Machine Learning
• Data Mining
• Computer Vision

Background:

• Spectral Clustering (SC) is widely used but faces challenges with suboptimal solutions from two-stage approaches.
• Existing SC methods struggle to maintain cluster balance and are computationally expensive for large datasets.

Purpose of the Study:

• To develop a novel Discrete and Balanced Spectral Clustering with Scalability (DBSC) model.
• To overcome the limitations of existing SC methods, including suboptimal solutions, lack of balance property, and poor scalability.

Main Methods:

• Integrated learning of continuous relaxation and discrete cluster indicator matrices into a single step.
• Incorporated an anchor-based strategy for improved scalability to large-scale datasets.
• Achieved soft-balanced clustering by maintaining approximately equal cluster sizes.

Main Results:

• The DBSC model demonstrated superior clustering and balance performance compared to existing methods.
• Achieved a significant 17.93% improvement in clustering accuracy on CMUPIE data over state-of-the-art methods.
• Successfully addressed issues of suboptimal solutions, cluster imbalance, and computational cost.

Conclusions:

• DBSC offers a more effective and efficient approach to spectral clustering.
• The model's integrated, balanced, and scalable design makes it suitable for real-world, large-scale applications.
• DBSC represents a significant advancement in spectral clustering techniques.