Shallow and ensemble deep randomized neural network for anomaly detection

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Anomaly detection, or one-class classification (OCC), is crucial for real-world applications.
• Traditional support vector machine-based OCC models struggle with large datasets and kernel sensitivity.
• Existing models often have limitations in capturing complex patterns due to single hidden layers.

Purpose of the Study:

• To propose novel deep learning models for enhanced anomaly detection.
• To overcome the limitations of traditional OCC methods, particularly with large-scale datasets.
• To improve the generalization, stability, and robustness of one-class classification models.

Main Methods:

• Introduction of the one-class random vector functional link (OC-RVFL) network, fusing linear and nonlinear patterns.
• Development of the one-class ensemble deep RVFL (OC-edRVFL) by integrating deep learning and ensemble learning with OC-RVFL.
• Utilizing a closed-form solution for efficient output weight computation and deriving generalization error bounds.

Main Results:

• The OC-edRVFL model demonstrates superior stability, robustness, and generalization compared to the OC-RVFL.
• Experiments on diverse datasets (artificial, UCI, NDC, MNIST) show OC-edRVFL outperforms baseline models.
• The proposed models exhibit high performance on datasets with up to 5 million samples.

Conclusions:

• The OC-edRVFL is a highly effective and scalable solution for anomaly detection.
• The novel deep ensemble approach significantly advances the capabilities of one-class classification.
• The models offer reduced training time and improved performance on large, complex datasets.