Spectral Clustering With Adaptive Neighbors for Deep Learning.

Area of Science:

• Computer Science
• Machine Learning
• Data Mining

Background:

• Spectral clustering is a powerful unsupervised learning technique.
• Traditional spectral clustering faces scalability challenges with large datasets due to computational costs in affinity matrix construction and eigen-decomposition.

Purpose of the Study:

• To develop a more efficient and effective spectral clustering algorithm for large-scale datasets.
• To address the limitations of traditional spectral clustering by optimizing affinity matrix construction and eigen-decomposition.

Main Methods:

• Proposes adaptive neighbor assignments for efficient affinity matrix construction, considering global data distribution.
• Introduces a deep learning framework with fully connected layers to learn a mapping function, replacing traditional eigen-decomposition.
• Evaluates the algorithm on both synthetic (toy) and real-world datasets.

Main Results:

• The proposed algorithm demonstrates significant improvements over existing clustering methods.
• Experimental results confirm the competitiveness and superiority of the new approach on various datasets.
• Achieves efficient handling of large-scale datasets, overcoming previous scalability issues.

Conclusions:

• The novel spectral clustering approach offers a scalable and effective solution for unsupervised learning on large datasets.
• The combination of adaptive neighbor assignments and deep learning provides a robust alternative to traditional spectral clustering methods.
• The algorithm shows strong performance and competitiveness across diverse experimental settings.