A Theoretical Analysis of Density Peaks Clustering and the Component-Wise Peak-Finding Algorithm

Area of Science:

• Data Science
• Machine Learning
• Statistical Modeling

Background:

• Density Peaks Clustering (DPC) is an unsupervised learning algorithm that identifies cluster centers based on local density and distance to higher-density points.
• While effective in practice, DPC's theoretical properties and robustness to noise require further investigation.

Purpose of the Study:

• To theoretically analyze the properties of Density Peaks Clustering.
• To propose a novel, robust clustering algorithm, Component-wise Peak-Finding (CPF), that addresses DPC's limitations.
• To evaluate the performance of CPF in various clustering tasks, including semi-supervised applications.

Main Methods:

• Theoretical analysis of DPC to establish its consistency in mode estimation and clustering accuracy under ideal conditions.
• Development of Component-wise Peak-Finding (CPF) algorithm, enhancing DPC by operating within density level sets and mitigating spurious maxima.
• Experimental validation of CPF using extensive datasets, comparing its performance against existing methods and demonstrating its effectiveness in semi-supervised computer vision tasks.

Main Results:

• Theoretical proof of DPC's consistency in estimating modes and high-probability data clustering.
• Demonstration that noise in density estimates can lead to erroneous modes and cluster assignments in DPC.
• CPF algorithm shows improved robustness to noise, automatically determines the correct number of clusters, and achieves exceptional performance in experiments.
• Semi-supervised CPF integrates clustering constraints for excellent performance in computer vision problems.

Conclusions:

• DPC provides a theoretically sound basis for mode detection and clustering but is sensitive to noise.
• CPF offers a significant advancement over DPC, providing a robust and automated solution for clustering.
• CPF's adaptability to semi-supervised learning enhances its utility for complex real-world applications like computer vision.