GMM Enhanced Anchor-Based Spectral Clustering for Large-Scale Data.

Area of Science:

• Machine Learning
• Data Mining
• Computational Statistics

Background:

• Traditional spectral clustering (SC) faces scalability challenges with large datasets.
• Existing anchor-based SC methods often overlook object membership heterogeneity, impacting anchor quality and clustering accuracy.

Purpose of the Study:

• To propose a novel Gaussian mixture model-enhanced spectral clustering (GMM-SC) approach for large-scale data.
• To address the limitations of existing anchor-based methods by accounting for object membership heterogeneity.

Main Methods:

• A two-stage divide-and-conquer strategy is employed.
• Stage 1: Gaussian mixture model (GMM) with expectation maximization (EM) algorithm categorizes objects into prior-consistent and prior-uncertain groups.
• Stage 2: Anchor-based SC is applied to prior-uncertain objects using anchors sampled from GMM components, followed by cluster alignment.

Main Results:

• The proposed GMM-SC approach significantly reduces computational complexity compared to traditional anchor-based SC.
• Experiments on large-scale datasets demonstrate the superiority of GMM-SC over existing state-of-the-art techniques.
• The method effectively handles object membership heterogeneity, leading to improved clustering accuracy.

Conclusions:

• GMM-SC offers an effective and efficient solution for large-scale spectral clustering.
• The novel approach enhances clustering accuracy by explicitly modeling object membership heterogeneity.
• This method provides a scalable and robust alternative for complex clustering tasks.