A novel bidirectional clustering algorithm based on local density.

Area of Science:

• Data Science
• Machine Learning
• Pattern Recognition

Background:

• Cluster analysis is widely applied, but determining the optimal number of clusters and identifying small, significant clusters remain challenging.
• Existing clustering algorithms often overlook small clusters, which are vital for uncovering extreme data characteristics.

Purpose of the Study:

• To propose a novel bidirectional clustering algorithm based on local density (BCALoD) that addresses the limitations of current methods.
• To enhance the detection of small clusters and automate the determination of the number of clusters.
• To introduce a denoising method robust to noise for improved clustering performance.

Main Methods:

• The BCALoD algorithm establishes data point connections based on local density.
• It automatically determines the number of clusters, reducing the need for parameter tuning.
• A denoising method is developed, assigning different cutoff distances and densities to clusters for enhanced robustness.

Main Results:

• BCALoD demonstrates increased sensitivity to small clusters compared to traditional algorithms.
• The algorithm automatically determines the number of clusters, minimizing manual parameter adjustments.
• The proposed denoising method further improves clustering performance, as validated on synthetic datasets and satellite imagery.

Conclusions:

• BCALoD offers an effective solution for cluster analysis, particularly for identifying small clusters and automating cluster number selection.
• The algorithm's robustness to noise and improved performance highlight its potential for diverse data analysis applications.
• BCALoD advances the field of clustering by providing a more sensitive and automated approach to data segmentation.