Unsupervised Feature Selection Using Bayesian Tucker Decomposition

Area of Science:

• Multivariate data analysis
• Machine learning
• Statistical modeling

Background:

• Traditional methods for feature selection and extraction often require supervision.
• Existing decomposition techniques may not adequately model uncertainty or complex data structures.
• There is a need for robust unsupervised methods applicable to diverse data types.

Purpose of the Study:

• To propose Bayesian Tucker Decomposition (BTuD) as a novel approach for unsupervised feature selection.
• To demonstrate the efficacy of BTuD across various synthetic and real-world datasets.
• To establish BTuD as a promising alternative to existing feature extraction techniques.

Main Methods:

• Development of a Bayesian Tucker Decomposition (BTuD) framework.
• Implementation of an algorithm for BTuD, consistent with higher-order orthogonal iteration.
• Application of BTuD for unsupervised feature selection on synthetic data, global coupled maps, and gene expression profiles.

Main Results:

• Successful application of BTuD for unsupervised feature selection on diverse datasets.
• Demonstration of BTuD's capability to handle complex data structures.
• Validation of BTuD's potential for feature extraction, aligning with prior tensor decomposition (TD) based methods.

Conclusions:

• Bayesian Tucker Decomposition (BTuD) is a promising new method for unsupervised feature selection.
• BTuD offers a robust approach for analyzing complex datasets without prior labels.
• BTuD-based feature extraction is expected to complement existing tensor decomposition techniques.