An Adaptive Density Distribution Clustering Method for Arbitrary-Shaped Datasets.

Area of Science:

• Data Science
• Machine Learning
• Computational Statistics

Background:

• Density peak clustering is effective for unlabeled data but struggles with accurate center selection.
• Existing methods' performance relies heavily on the precise identification of cluster centers.

Purpose of the Study:

• To develop an adaptive density distribution clustering (ADDC) method to overcome challenges in selecting cluster centers.
• To introduce a robust and decentralized clustering approach for complex datasets.

Main Methods:

• Constructing an undirected neighborhood graph using a novel neighbor degree definition for decentralized allocation.
• Introducing componentwise local density and a new criterion for density peak selection to guide cluster number determination.
• Formulating criterion-based decomposition and fusion strategies using the neighborhood graph and density peaks to identify and refine clusters.

Main Results:

• ADDC demonstrates superior performance compared to five classical and seven state-of-the-art density-based clustering methods.
• The method effectively identifies clusters with multiple peaks and detects low-density clusters lacking distinct peaks.
• Experiments on real and synthetic datasets validate the robustness and effectiveness of ADDC.

Conclusions:

• ADDC offers a significant advancement in density-based clustering, particularly in handling complex data structures.
• The proposed method provides a more accurate and reliable approach to cluster center selection and identification.
• ADDC enhances knowledge discovery from unlabeled datasets through improved clustering performance.