Detectability Driven Recommendation of Anomaly Detection Models for Time-Series Data.

Area of Science:

• Time-series analysis
• Machine learning
• Data mining

Background:

• Anomaly detection in time-series data is crucial for practical applications.
• Existing deep learning models require offline training and face limitations with multiple models due to computational costs.
• Current model recommendation methods are often inefficient or perform poorly with real-world data challenges like missing labels and heterogeneous characteristics.

Purpose of the Study:

• To propose a novel and efficient recommendation method for selecting appropriate anomaly detection models.
• To address the limitations of existing recommendation systems in terms of time efficiency and performance.
• To enable effective online anomaly detection with limited computational resources.

Main Methods:

• Introduced a model recommendation framework based on the concept of 'detectability'.
• Defined detectability using a fine-grained strategy for comparing data characteristics.
• Developed an efficient algorithm for computing detectability and making model recommendations.

Main Results:

• Extensive experiments were conducted on real time-series data.
• The proposed method demonstrated effectiveness in selecting anomaly detection models.
• The approach proved to be computationally efficient.

Conclusions:

• The novel recommendation method effectively addresses the challenges of selecting anomaly detection models for time-series data.
• The proposed framework and algorithm offer an efficient solution for online anomaly detection applications.
• This work contributes to improving the practical deployment of anomaly detection systems.