A self-organizing network for hyperellipsoidal clustering (HEC).

Area of Science:

• Machine Learning
• Data Mining
• Pattern Recognition

Background:

• Traditional clustering algorithms often assume spherical cluster shapes, limiting their effectiveness with real-world data.
• Estimating Mahalanobis distance can be problematic with limited data points relative to feature dimensionality.

Purpose of the Study:

• To propose a self-organizing network for hyperellipsoidal clustering (HEC).
• To address limitations in Mahalanobis distance estimation for clusters with few data points.
• To develop a clustering method that balances hyperspherical and hyperellipsoidal cluster shapes.

Main Methods:

• A two-layer network utilizing principal component analysis subnetworks for shape estimation.
• Competitive learning incorporating cluster shape information.
• Partitional clustering with a regularized Mahalanobis distance.
• Kolmogorov-Smirnov test significance level as a compactness measure.

Main Results:

• The HEC network demonstrated significant improvements over K-means with Euclidean distance on artificial and real datasets.
• The method effectively handles challenges in Mahalanobis distance estimation.
• The network prevents the formation of unusually large or small clusters.
• Experiments revealed the frequent occurrence of hyperellipsoidal clusters in practical scenarios.

Conclusions:

• The proposed HEC network offers a robust solution for clustering data with hyperellipsoidal structures.
• This approach provides superior clustering performance compared to Euclidean-based K-means.
• The findings highlight the prevalence of non-spherical cluster shapes in real-world datasets.