DKCDC: A clustering algorithm focusing on genuine boundary search for regional division

Area of Science:

• Data Mining
• Machine Learning
• Pattern Recognition

Background:

• Existing clustering algorithms often fail to produce well-defined boundaries.
• Difficulty in achieving genuine and reliable cluster boundaries hinders regional division.

Purpose of the Study:

• To propose a novel clustering algorithm, DKCDC, for enhanced regional division.
• To address the limitations of existing methods in boundary detection and validation.

Main Methods:

• DKCDC integrates Direction Centrality (CDC) with the Distance of K-nearest-neighbor.
• A fusion strategy combines voting and distance metrics to differentiate true from false boundaries.
• Noise points within boundaries are identified and processed for improved accuracy.

Main Results:

• DKCDC achieves well-defined regional boundaries.
• The algorithm demonstrated a significant improvement in silhouette coefficient by at least 4.88% compared to CDC, K-Means, DBSCAN, OPTICS, and HDBSCAN.
• Experiments on synthetic and UCI datasets validated the effectiveness of DKCDC.

Conclusions:

• DKCDC offers a robust solution for clustering-based regional division with enhanced boundary demarcation.
• The proposed fusion strategy is key to distinguishing genuine boundaries and improving clustering performance.
• DKCDC shows broad potential for applications requiring precise regional identification.