Hierarchical One-Class Classifier With Within-Class Scatter-Based Autoencoders.

Area of Science:

• Machine Learning
• Deep Neural Networks
• Representation Learning

Background:

• Autoencoding is crucial for representation learning in deep neural networks (DNNs).
• Extreme Learning Machine-based Autoencoders (ELM-AEs) offer fast learning but may produce suboptimal features due to random parameter initialization.
• A need exists for autoencoders that generate more meaningful encoded features.

Purpose of the Study:

• To propose a Within-Class Scatter Information constraint-based Autoencoder (WSI-AE).
• To enhance the quality of encoded features by minimizing reconstruction error and within-class scatter.
• To integrate WSI-AEs into a one-class classification (OCC) algorithm.

Main Methods:

• Developed a WSI-AE model that constrains within-class scatter of encoded features.
• Constructed stacked WSI-AEs for a one-class classification (OCC) task.
• Employed a hierarchical regularized least-squared method for the OCC algorithm.

Main Results:

• The proposed WSI-AE effectively minimizes reconstruction error and within-class scatter.
• Stacked WSI-AEs demonstrated superior performance in one-class classification tasks.
• Experimental results showed advantages over existing state-of-the-art autoencoders and OCC algorithms.

Conclusions:

• The WSI-AE approach yields more meaningful encoded features compared to standard ELM-AEs.
• The developed WSI-AE-based OCC algorithm is effective on benchmark datasets.
• This method offers a promising direction for representation learning and classification.