Detecting Regions of Maximal Divergence for Spatio-Temporal Anomaly Detection

Area of Science:

• Data Science
• Machine Learning
• Time Series Analysis

Background:

• Anomaly detection is crucial in fields like fraud detection, climate analysis, and healthcare.
• Existing methods often focus on isolated data points, not coherent regions.
• Handling large-scale, multivariate spatio-temporal data presents significant challenges.

Purpose of the Study:

• To present a novel algorithm for detecting anomalous regions in multivariate spatio-temporal time-series data.
• To enable unsupervised detection of coherent spatial and temporal anomalies.
• To provide a scalable solution for analyzing large datasets.

Main Methods:

• Developed the "Maximally Divergent Intervals" (MDI) framework.
• Defined an unbiased Kullback-Leibler divergence for ranking regions of varying sizes.
• Implemented an interval proposal technique for efficient large-scale data processing.

Main Results:

• The MDI framework successfully detects anomalous regions and time intervals.
• The method is effective in unsupervised anomaly detection.
• Experiments show wide applicability across diverse data domains.

Conclusions:

• The proposed MDI framework is a valuable tool for identifying significant events in large, complex datasets.
• The algorithm demonstrates broad applicability in climate analysis, video surveillance, and text forensics.
• This method advances the field of spat-temporal anomaly detection.