MultiverseAD: Enhancing spatial-temporal synchronous attention networks with causal knowledge for multivariate time...

Area of Science:

• Computer Science
• Artificial Intelligence
• Data Science

Background:

• Multivariate time series anomaly detection is vital for system reliability.
• Existing deep learning methods often fail to capture complex spatial-temporal interactions.

Purpose of the Study:

• To introduce MultiverseAD, a novel spatial-temporal synchronous attention network.
• To enhance anomaly detection in multivariate time series by incorporating causal knowledge.

Main Methods:

• Developed MultiverseAD, combining a dynamic spatial-temporal synchronous attention network with a static spatial-temporal causal graph.
• Employed sliding graph attention for local and long-term dependencies.
• Integrated causal graph to encode static causal relationships.

Main Results:

• MultiverseAD demonstrated superior performance across eight public datasets.
• Consistently outperformed twelve state-of-the-art anomaly detection models.
• Ablation studies confirmed the effectiveness of the causal graph and synchronous attention mechanisms.

Conclusions:

• MultiverseAD offers a significant advancement in multivariate time series anomaly detection.
• The integration of causal knowledge and spatial-temporal synchronous attention is key to improved performance.