Improved initialisation of model-based clustering using Gaussian hierarchical partitions.

Area of Science:

• Statistics
• Machine Learning
• Data Mining

Background:

• The Expectation-Maximization (EM) algorithm is widely used for model-based clustering.
• Initialization of the EM algorithm significantly impacts clustering results, potentially leading to suboptimal solutions (local maxima).
• Current methods, like model-based agglomerative hierarchical clustering in the mclust R package, offer convenience but can yield poor initial partitions.

Purpose of the Study:

• To address the issue of poor initializations in EM-based model-based clustering.
• To propose simple, fast, and effective refinement techniques for initial clustering partitions.
• To enhance the reliability and accuracy of clustering outcomes by improving EM algorithm initialization.

Main Methods:

• Investigated the impact of different initialization strategies on EM algorithm convergence.
• Developed and applied data transformation techniques to refine initial clustering partitions.
• Utilized model-based agglomerative hierarchical clustering as a baseline for comparison.

Main Results:

• Proposed data transformation methods provide simple and fast refinements for initial clustering partitions.
• These refinements can help mitigate the convergence to local maxima of the likelihood function.
• Illustrative examples demonstrate the practical utility of the proposed techniques.

Conclusions:

• Data transformation-based refinements offer a valuable improvement over standard initialization in model-based clustering.
• These methods enhance the robustness of the EM algorithm by guiding it towards better solutions.
• The proposed techniques are computationally efficient and easy to implement, improving clustering performance.