Model-based clustering by probabilistic self-organizing maps.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Mining

Background:

• Self-organizing maps (SOM) are effective for topology preservation in data clustering.
• Existing methods may not fully leverage probabilistic models for enhanced clustering.
• Integrating probabilistic approaches can improve the robustness and interpretability of SOMs.

Purpose of the Study:

• To develop a probabilistic self-organizing map (PbSOM) model for model-based data clustering.
• To extend Kohonen's SOM by incorporating multivariate Gaussian distributions for reference vectors.
• To introduce novel expectation-maximization (EM)-type algorithms for learning the PbSOM model.

Main Methods:

• Developed a coupling-likelihood mixture model for PbSOM.
• Derived three EM-type algorithms: SOCEM, SOEM, and SODAEM.
• Utilized classification EM (CEM) and standard EM algorithms, along with deterministic annealing (DA).

Main Results:

• The proposed PbSOM learning algorithms achieve effective data clustering.
• Performance is comparable to the deterministic annealing EM (DAEM) approach.
• The topology-preserving property of the data clusters is successfully maintained.

Conclusions:

• The developed PbSOM model and its associated learning algorithms offer a robust approach to topology-preserving data clustering.
• These methods provide a probabilistic framework that enhances traditional SOMs.
• The algorithms demonstrate competitive performance and preserve essential data structures.