Novel unsupervised feature filtering of biological data

Area of Science:

• Computational Biology
• Data Science
• Bioinformatics

Background:

• Unsupervised feature selection is crucial for large, noisy datasets.
• Existing unsupervised methods are limited.
• Novel methods are needed to identify informative feature subsets.

Purpose of the Study:

• To propose a novel unsupervised feature selection criterion based on SVD-entropy.
• To evaluate the effectiveness of this new method in identifying informative features.
• To compare the proposed method against existing techniques.

Main Methods:

• Developed a novel unsupervised criterion using SVD-entropy.
• Calculated contribution to entropy (CE) on a leave-one-out basis.
• Implemented four variations: simple ranking (SR), forward selection (FS1, FS2), and backward elimination (BE).

Main Results:

• Applied methods to benchmark datasets and evaluated clustering success using Jaccard scores.
• Feature filtering by CE outperformed variance and gene-shaving methods.
• Selected feature sets achieved superior clustering performance compared to using all features in some cases.
• Identified an optimal feature set size, typically a small percentage of total features.
• Selected genes showed significant Gene Ontology (GO) enrichment in relevant cellular processes.

Conclusions:

• The SVD-entropy based feature selection method is effective for unsupervised learning.
• This approach can identify a small subset of highly informative features.
• The selected features are biologically relevant and improve data clustering performance.