A novel autoencoder approach to feature extraction with linear separability for high-dimensional data.

Area of Science:

• Machine Learning
• Data Science
• Dimensionality Reduction

Background:

• High-dimensional data presents challenges for feature extraction due to sparse distributions.
• Locating features in subspaces is difficult in high-dimensional spaces.

Purpose of the Study:

• To propose a novel autoencoder method for effective feature extraction from high-dimensional data.
• To improve the accuracy and linear separability of extracted features.

Main Methods:

• A novel autoencoder method utilizing Mahalanobis distance metric of rescaling transformation.
• Reducing the distribution difference between reconstructed and original data.

Main Results:

• The proposed method achieves superior accuracy in feature extraction compared to state-of-the-art techniques.
• Extracted features demonstrate enhanced linear separability.
• Distance metric-based methods are more effective than feature selection for linear separability in high-dimensional data.

Conclusions:

• The Mahalanobis distance-based autoencoder is effective for feature extraction in high-dimensional spaces.
• Distance metric methods offer advantages over feature selection methods for extracting linearly separable features.
• Feature similarity evaluation is more suitable than feature importance for high-dimensional data feature extraction.