An Unsupervised Fusion Strategy for Anomaly Detection via Chebyshev Graph Convolution and a Modified Adversarial...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Anomalies in datasets are deviations from the norm, often indicating errors or unusual events.
• Traditional anomaly detection is slow and prone to errors, hindering system and network performance optimization.
• Automatic anomaly recognition is crucial for reducing defect costs and enhancing operational efficiency.

Purpose of the Study:

• To propose an unsupervised deep network for accurate time series anomaly prediction.
• To develop an efficient method for detecting abnormal samples within fluctuating time series data.
• To reduce the cost and errors associated with traditional anomaly detection techniques.

Main Methods:

• An unsupervised deep network was developed to predict temporal information.
• A graph of neighboring fluctuations was constructed by analyzing correlations between time series samples.
• Chebyshev graph convolution layers were applied to extracted temporal features, feeding into a modified generative adversarial network (GAN).

Main Results:

• The proposed Chebyshev graph-based modified adversarial network (Cheb-MA) demonstrated strong performance.
• The model achieved an average F1-score of 82.09% on the Numenta dataset.
• The results show significant promise compared to existing research in anomaly detection.

Conclusions:

• The fusion of generative adversarial networks (GANs) and Chebyshev graphs offers an effective approach for anomaly detection.
• The Cheb-MA model provides an efficient and accurate unsupervised method for identifying anomalies in time series.
• This technique has the potential to optimize system performance and reduce operational costs through automated defect detection.