Enhancing Unsupervised Anomaly Detection With Score-Guided Network.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Anomaly detection is vital in healthcare and finance, but limited labels necessitate unsupervised methods.
• Existing unsupervised methods struggle to differentiate mixed normal/abnormal data and effectively maximize score disparities.
• Representation learning is key, but defining metrics to separate data in hypothesis space remains challenging.

Purpose of the Study:

• To propose a novel scoring network with score-guided regularization for enhanced anomaly detection.
• To improve the learning of informative representations, particularly for data in transition fields.
• To offer a plug-in component enhancing existing unsupervised representation learning models.

Main Methods:

• Developed a scoring network with score-guided regularization to enlarge anomaly score disparities.
• Integrated the scoring network into an autoencoder and four state-of-the-art unsupervised representation learning models (SG-Models).
• Evaluated the effectiveness and transferability of the proposed approach on synthetic and real-world datasets.

Main Results:

• The proposed score-guided strategy enables gradual learning of more informative representations.
• SG-Models demonstrated state-of-the-art performance across diverse datasets.
• The scoring network functions effectively as a plug-in component for various deep unsupervised anomaly detection models.

Conclusions:

• The novel scoring network and score-guided regularization significantly enhance unsupervised anomaly detection capabilities.
• SG-Models offer a versatile and effective solution for identifying anomalies in complex systems.
• The approach shows strong effectiveness and transferability, outperforming existing methods.