An Online Method to Detect Urban Computing Outliers via Higher-Order Singular Value Decomposition

Area of Science:

• Data Science
• Urban Computing
• Signal Processing

Background:

• Urban spaces generate vast, multiway data streams from sensors.
• Detecting outliers in real-time urban data is crucial for system monitoring and anomaly identification.
• Traditional offline methods may not efficiently handle the dynamic nature of streaming urban data.

Purpose of the Study:

• To propose and evaluate an online method for outlier detection in multiway urban space data.
• To leverage higher-order singular value tensor decomposition for real-time anomaly identification.
• To compare the performance of the proposed online method against traditional offline approaches.

Main Methods:

• Developed a two-step online outlier detection method: offline modeling and online modeling.
• Utilized higher-order singular value tensor decomposition (HOSVD) for data analysis.
• Applied the method to real-time streaming sensor data from three Finnish cities: Helsinki, Tuusula, and Lohja.

Main Results:

• The online outlier detection method achieved higher accuracy than the offline approach.
• Accuracy gains ranged from 8.5% to 10% compared to the offline method.
• Real-time monitoring using a sliding window proved more effective for outlier detection.

Conclusions:

• The proposed online method offers a more accurate and efficient solution for outlier detection in urban sensor data.
• Real-time processing via sliding window enhances the adequacy of outlier detection.
• HOSVD is a powerful tool for analyzing the multiway nature of urban data for anomaly detection.