An online forecasting-based fine-tuning pipeline for time-series anomaly prediction.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Time-series anomaly detection is crucial but limited to complete data.
• Existing methods require ground truth, hindering prediction of future anomalies.
• A gap exists in predicting anomalies without prior knowledge of ground truth.

Purpose of the Study:

• Introduce Time-Series Anomaly Prediction (TSAP) for forecasting anomalies.
• Develop a method to predict anomaly occurrence and progression without ground truth.
• Address limitations of current anomaly detection and forecasting techniques.

Main Methods:

• Propose an exemplar-based pre-training and fine-tuning pipeline.
• Utilize online time-series forecasting techniques for anomaly prediction.
• Employ a three-step online process: prediction/detection, motif search, and exemplar fine-tuning.

Main Results:

• Achieve up to 53.8% improvement in anomaly detection F1 score.
• Enhance time-series forecasting accuracy during anomalies by up to 82.4% (MSE).
• Improve post-anomaly time-series forecasting accuracy by up to 49.1% (MSE).

Conclusions:

• The proposed TSAP method effectively addresses the challenge of predicting future anomalies.
• Demonstrate superior performance compared to state-of-the-art methods on real-world and synthetic data.
• Highlight the method's capability for tasks currently unaddressed by existing anomaly detection or forecasting techniques.