Fractal Autoencoders for Feature Selection.

Area of Science:

• Machine Learning
• Bioinformatics
• Data Science

Background:

• High-dimensional data presents challenges in analysis and interpretation.
• Effective feature selection is crucial for dimensionality reduction and uncovering meaningful patterns.
• Unsupervised methods are needed to identify informative features without prior labels.

Purpose of the Study:

• To introduce an innovative unsupervised feature selection framework, Fractal Autoencoders (FAE).
• To enable efficient identification of informative features for both global representability and local diversity.
• To demonstrate the effectiveness and extensibility of the FAE framework.

Main Methods:

• Developed Fractal Autoencoders (FAE), an extension of autoencoders.
• Integrated a one-to-one scoring layer and a sub-neural network for unsupervised feature selection.
• Trained the neural network to pinpoint informative features by exploring data representability and diversity.

Main Results:

• FAE achieved state-of-the-art performance on unsupervised feature selection across fourteen diverse datasets.
• Demonstrated superiority over existing contemporary methods, especially on very high-dimensional data.
• Showcased significant advantages in gene expression data exploration, reducing measurement costs by approximately 15% compared to L1000 landmark genes.

Conclusions:

• FAE provides a powerful and concise architecture for effective unsupervised feature selection.
• The framework offers substantial benefits for high-dimensional data analysis and biological data exploration.
• FAE is a versatile and extensible framework with practical applications.