Causality in scale space as an approach to change detection

Area of Science:

• Time series analysis
• Statistical modeling
• Data visualization

Background:

• Kernel density estimation and kernel regression are valuable for data structure assessment.
• Detecting changes in time series is crucial for live surveillance systems.

Purpose of the Study:

• To develop a method for early change point detection in time series data.
• To enable reliable change point identification without prior scale knowledge.
• To provide retrospective interval estimates for change point timing.

Main Methods:

• Definition of a temporal scale space using bandwidth and temporal variables.
• Identification of significance regions via hypothesis tests for significant gradients.
• Imposition of causality using left-bounded kernels and forward shifting.
• Application of kernel density estimation and kernel regression.

Main Results:

• The method enables early detection of changes in time series data.
• Warning delays are comparable to standard techniques, with advantages in scale-agnostic detection.
• Reliable detection of change points is achieved even with limited scale information.
• Retrospective reliable interval estimates of change point timing are obtained.

Conclusions:

• The proposed technique is applicable to diverse data sources without tailoring.
• It offers a robust approach for live surveillance systems, including disease outbreak detection and patient monitoring.
• The method provides a valuable alternative for change point analysis, especially when the relevant scale is unknown.