A simple method for unsupervised anomaly detection: An application to Web time series data.

Area of Science:

• Data Science
• Machine Learning
• Time Series Analysis

Background:

• Anomaly detection in unlabeled time series data is challenging.
• Existing methods may lack tractability for non-technical users.

Purpose of the Study:

• To propose a simple and tractable anomaly detection method for unlabeled time series data.
• To evaluate the method's performance on benchmark and real-world datasets.

Main Methods:

• Utilizes density ratio estimation based on a state space model.
• Employs a detection rule based on the ratio of log-likelihoods from a dynamic linear model and a NULL model.

Main Results:

• Achieves comparable or superior performance on Yahoo S5 and Numenta Anomaly Benchmark datasets.
• Successfully applied to unlabeled web time series data (e-commerce page views, session duration).
• Identified that increased page views from newsletters are less likely to lead to insurance contracts.

Conclusions:

• Incorporating time series specific information is crucial for effective anomaly detection.
• The proposed method is applicable to unlabeled real-world data.
• Simultaneous monitoring of multiple time series is recommended.