An Improved Density Peak Clustering Algorithm for Multi-Density Data.

Area of Science:

• Data Science
• Machine Learning
• Artificial Intelligence

Background:

• Density Peak Clustering (DPC) is an iterative-free clustering algorithm effective for identifying cluster centers.
• Standard DPC struggles with multi-density datasets due to a single parameter limitation and subjective cluster center selection.
• Existing methods often yield suboptimal clustering results and lack convincing cluster center determination for complex data.

Purpose of the Study:

• To improve the Density Peak Clustering algorithm for enhanced performance on multi-density datasets.
• To develop an automated method for parameter selection and cluster center identification in DPC.
• To address the limitations of subjective decision-making in traditional DPC implementations.

Main Methods:

• Introduced K-nearest neighbor analysis to determine optimal parameter 'd' by identifying global bifurcation points for density division.
• Developed a 'gamma map' and calculated average gamma height differences to automatically detect cluster centers and their quantity.
• Applied the improved DPC algorithm to divided datasets, followed by cluster fusion rules for result refinement.

Main Results:

• The enhanced DPC algorithm, termed F-DPC, successfully clusters multi-density data by automatically selecting parameters and cluster centers.
• Experimental results on simulated and UCI datasets demonstrate superior clustering effect and quality compared to standard DPC.
• The algorithm shows robustness across varying numbers of samples and dataset complexities.

Conclusions:

• The proposed F-DPC algorithm effectively overcomes the limitations of traditional DPC for multi-density data.
• Automated parameter and cluster center selection significantly improves clustering objectivity and accuracy.
• F-DPC offers a more reliable and efficient clustering solution for diverse real-world applications.