Recurrent Reconstructive Network for Sequential Anomaly Detection.

Area of Science:

• Computer Science
• Machine Learning
• Data Mining

Background:

• Anomaly detection is crucial for identifying rare, deviating data points.
• Imbalanced datasets are common in anomaly detection, often addressed by one-class classification.
• Recurrent Autoencoders (RAE) show promise for sequential anomaly detection but struggle with long-term dependencies and fixed-length inputs.

Purpose of the Study:

• To propose a novel Recurrent Reconstructive Network (RRN) to overcome RAE limitations in sequential anomaly detection.
• To enhance anomaly detection for streaming data with variable sequence lengths.
• To improve the accuracy and robustness of anomaly detection models.

Main Methods:

• Developed a Recurrent Reconstructive Network (RRN) building upon RAE architecture.
• Incorporated a self-attention mechanism to manage varying input sequence lengths.
• Implemented hidden state forcing and skip transitions with attention gates to improve reconstruction and handle sequence variations.

Main Results:

• The proposed RRN effectively manages input sequences of varying lengths.
• The self-attention mechanism and hidden state forcing enhance the model's ability to process sequential data.
• Skip transitions with attention gates significantly improve reconstruction performance.
• Comprehensive experiments on four datasets validated the superior performance of RRN over conventional methods.

Conclusions:

• The Recurrent Reconstructive Network (RRN) offers a significant advancement in sequential anomaly detection.
• RRN's novel functionalities address key limitations of existing RAE models.
• The model demonstrates high efficacy in detecting anomalies in streaming data with imbalanced distributions.