Clustering by measuring local direction centrality for data with heterogeneous density and weak connectivity

Area of Science:

• Data Science
• Machine Learning
• Bioinformatics

Background:

• Clustering algorithms are vital for pattern discovery across diverse scientific fields.
• Existing methods struggle with datasets exhibiting weak connectivity and varying densities.
• Limitations in current clustering approaches necessitate novel algorithms for complex data structures.

Purpose of the Study:

• To introduce a novel clustering algorithm, Centrality-based Directionality (CDC).
• To address the limitations of traditional clustering methods in handling heterogeneous data densities and weak connectivity.
• To enhance the accuracy and robustness of pattern discovery in complex datasets.

Main Methods:

• Developed a boundary-seeking clustering algorithm utilizing local Direction Centrality (CDC).
• Employed a density-independent metric based on K-nearest neighbors (KNNs) distribution.
• Utilized boundary points to form 'cages' that isolate internal points, preventing cross-cluster connections.

Main Results:

• Demonstrated CDC's effectiveness in detecting complex clusters in synthetic datasets.
• Successfully identified cell types from single-cell RNA sequencing (scRNA-seq) and mass cytometry (CyTOF) data.
• Validated CDC's performance on various real-world benchmarks, including speaker recognition.

Conclusions:

• CDC offers a robust solution for clustering challenges posed by weak connectivity and heterogeneous densities.
• The algorithm's ability to distinguish boundary and internal points enhances cluster separation.
• CDC shows broad applicability across various domains, from bioinformatics to voice analysis.