Robust k-Means-Type Clustering for Noisy Data.

Area of Science:

• Machine Learning
• Data Mining
• Statistical Modeling

Background:

• Data clustering is essential for grouping similar data points.
• Real-world datasets often contain noise, challenging traditional clustering algorithms like k-means.
• Gaussian Mixture Models (GMM) provide a probabilistic foundation for k-means.

Purpose of the Study:

• To develop a robust k-means-type clustering algorithm resistant to noise.
• To leverage the properties of the t-distribution for enhanced clustering performance.
• To offer a simpler yet effective alternative to full t-mixture models.

Main Methods:

• Proposing a novel k-means-type clustering algorithm named KMTD.
• Assuming data points are generated from a multivariate t-mixture model (TMM).
• Utilizing the fatter tails of the t-distribution to mitigate noise impact.

Main Results:

• KMTD demonstrates increased robustness to noisy data compared to Gaussian-based methods.
• The algorithm achieves higher accuracy in most cases across synthetic and real datasets.
• KMTD exhibits faster execution times than other sophisticated clustering algorithms.

Conclusions:

• The proposed KMTD algorithm provides a robust and efficient solution for data clustering in the presence of noise.
• The t-distribution's properties make it suitable for modeling noisy data in clustering tasks.
• KMTD offers a practical balance between simplicity and performance for data analysis.