Large Scale Spectral Clustering Via Landmark-Based Sparse Representation.

Area of Science:

• Machine Learning
• Data Mining
• Computational Science

Background:

• Spectral clustering is a popular data clustering technique.
• Large-scale spectral clustering faces computational challenges (O(n^3)) and existing acceleration methods often degrade performance by losing data information.

Purpose of the Study:

• To propose a novel, efficient, and scalable spectral clustering algorithm for large-scale problems.
• To address the performance degradation associated with existing accelerated spectral clustering methods.

Main Methods:

• Developed a landmark-based spectral clustering approach.
• Selected a subset of data points (landmarks) to represent the entire dataset.
• Represented original data points as sparse linear combinations of landmarks.
• Computed spectral embeddings efficiently using the landmark-based representation.

Main Results:

• The proposed algorithm achieves linear scalability with the problem size.
• Demonstrated effectiveness and efficiency compared to state-of-the-art methods through extensive experiments.
• Preserved data information, avoiding performance degradation seen in other accelerated methods.

Conclusions:

• Landmark-based spectral clustering offers an effective and efficient solution for large-scale clustering.
• The method scales linearly, making it suitable for big data applications.
• This approach provides a promising alternative to existing spectral clustering techniques for large datasets.