Spontaneous clustering via minimum γ-divergence.

Area of Science:

• Data Science
• Machine Learning
• Statistics

Background:

• Traditional clustering algorithms (K-means, fuzzy c-means, model-based) require users to pre-define the number of clusters.
• This pre-specification can be a significant limitation, as the optimal number of clusters often reflects the underlying data structure and may not be known beforehand.

Purpose of the Study:

• To introduce a novel clustering method, termed spontaneous clustering, that automatically determines the number of clusters.
• To demonstrate the advantage of spontaneous clustering over existing methods by automatically detecting data structure.

Main Methods:

• Proposes a new clustering method based on the local minimization of the gamma-divergence.
• Identifies cluster centers by detecting local minimum points of the gamma-divergence.
• Derives a necessary and sufficient condition for the existence of local minimum points in a simplified setting.

Main Results:

• The proposed spontaneous clustering method automatically identifies the number of clusters inherent in the data.
• Demonstrates the method's effectiveness through applications to both simulated and real-world datasets.
• Compares the performance of spontaneous clustering against established clustering techniques.

Conclusions:

• Spontaneous clustering offers a significant advantage by eliminating the need to pre-define the number of clusters.
• The method effectively captures the intrinsic structure of data by identifying local minima of the gamma-divergence.
• This approach provides a more data-driven and automated solution for cluster analysis.