Feature Selection for Unsupervised Machine Learning.

Area of Science:

• Computer Science
• Machine Learning
• Data Mining

Background:

• Feature selection is less developed for unsupervised machine learning (ML) compared to supervised ML.
• Clustering algorithms like Gaussian Mixture Models (GMM) and k-means often utilize all available features, potentially impacting performance.
• There is a need for improved feature selection techniques in unsupervised learning to enhance clustering accuracy and efficiency.

Purpose of the Study:

• To propose a stepwise feature selection approach for unsupervised clustering methods.
• To adapt and improve Gaussian Mixture Models (GMM) and k-means algorithms through feature subset selection.
• To evaluate the performance of the proposed method against existing approaches using simulations and real-world data.

Main Methods:

• A stepwise feature selection strategy was developed for clustering.
• The proposed method selects an optimal subset of features for GMM and k-means.
• Gaussian Mixture Models (GMM) and k-means were potentially modified with improved initializations.

Main Results:

• The proposed feature selection method demonstrated superior accuracy and computational efficiency compared to using all features.
• Experiments using Monte Carlo simulations validated the effectiveness of the approach.
• A real-world dataset analysis confirmed the findings from the simulations, highlighting practical applicability.

Conclusions:

• The developed stepwise feature selection approach enhances the performance of GMM and k-means clustering.
• Selecting relevant features improves computational efficiency and accuracy in unsupervised ML.
• The findings suggest a practical improvement for applying GMM and k-means to complex datasets.