Hierarchical One-Class Model With Subnetwork for Representation Learning and Outlier Detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Multilayer one-class classification (OCC) frameworks are crucial for anomaly and outlier detection.
• Existing algorithms often struggle with loosely connected feature coding, hindering discriminative representation in the latent space.

Purpose of the Study:

• To propose two novel OCC frameworks, OC-SNN and MCOC-SNN, designed to enhance feature representation and classification performance.
• To address the limitations of existing multilayer OCC methods in generating discriminative latent spaces.

Main Methods:

• The study introduces the OCC structure using subnetwork neural network (OC-SNN) and maximum correntropy-based OC-SNN (MCOC-SNN).
• These models utilize subnetworks for discriminative latent space construction and function as one-step learning networks.
• MCOC-SNN employs the maximum correntropy criterion (MCC) for feature encoding, differing from traditional mean square error (MSE) approaches.

Main Results:

• Both OC-SNN and MCOC-SNN demonstrated superior performance compared to existing multilayer OCC models.
• Experiments were conducted on a new OCC dataset, CO-Mask, across a wide range of training samples.
• The proposed methods effectively generated highly discriminative representations for improved classification.

Conclusions:

• The novel OC-SNN and MCOC-SNN frameworks offer significant improvements in one-class classification for anomaly detection.
• The use of subnetworks and MCC provides a more effective approach to learning discriminative features.
• The developed CO-Mask dataset and open-source code facilitate further research and reproducibility.