A Flexible EM-Like Clustering Algorithm for Noisy Data.

Area of Science:

• Machine Learning
• Statistical Modeling
• Data Mining

Background:

• The Expectation-Maximisation (EM) algorithm for Gaussian mixture models struggles with non-Gaussian data, outliers, and noise.
• Existing clustering methods often lack robustness in complex, real-world scenarios.

Purpose of the Study:

• To introduce a novel clustering algorithm, the Flexible EM-like Clustering Algorithm (FEMCA).
• To enhance clustering robustness by accommodating heavier tail distributions, noise, and outliers.

Main Methods:

• FEMCA employs an EM-like procedure estimating cluster centers and covariances.
• A semi-parametric approach estimates a unique scale parameter per data point.
• The algorithm is analyzed for independent, non-identically distributed samples of elliptical distributions.

Main Results:

• FEMCA demonstrates robustness to heavier tail distributions, noise, and outliers.
• The algorithm exhibits important distribution-free properties.
• Convergence and accuracy are analyzed using synthetic data.

Conclusions:

• FEMCA significantly outperforms classical unsupervised methods like k-means and standard EM.
• The algorithm shows superior performance on benchmark real-world datasets (MNIST, NORB, 20newsgroups).
• FEMCA offers a more efficient and flexible alternative for complex clustering tasks.