DECA: A Discrete-Valued Data Clustering Algorithm.

Area of Science:

• Data Science
• Machine Learning
• Statistical Analysis

Background:

• Analyzing unordered discrete-valued data presents challenges for existing clustering algorithms.
• Current methods may be sensitive to the choice of distance measures, impacting solution reliability.

Purpose of the Study:

• To introduce a new clustering algorithm designed for unordered discrete-valued data.
• To enhance the robustness of clustering by minimizing sensitivity to arbitrary distance measures.

Main Methods:

• The algorithm employs a two-phase approach: cluster initiation and sample regrouping.
• Cluster initiation utilizes data-directed valley detection with optimal second-order product approximation and a discrete distance measure.
• Sample regrouping iteratively applies the Bayes' decision rule based on subgroup discrete distributions.

Main Results:

• The proposed algorithm was evaluated on four simulated datasets and one clinical dataset.
• Performance was compared against the established decision-directed algorithm.
• Experiments confirmed the algorithm's validity, operational feasibility, and superior performance.

Conclusions:

• The novel clustering algorithm provides a robust and effective solution for discrete-valued data analysis.
• Its probability-driven approach offers advantages over methods relying heavily on distance measures.
• The algorithm shows significant promise for applications in data mining and statistical analysis.