Robust Multiple Change Point Detection for Streaming Data
Saved in:
| Title: | Robust Multiple Change Point Detection for Streaming Data |
|---|---|
| Authors: | Pandohie, Randyll |
| Committee Members: | Maboudou, Edgard |
| Summary: | The growing prevalence of high-frequency, high-dimensional data streams in domains such as finance, cybersecurity, and industrial monitoring has intensified the demand for real-time multiple change point detection methods. These methods are expected to identify distributional shifts as they occur while also determining the number and locations of change points—even in the presence of noise, outliers, and limited labeled data. Traditional batch-based approaches and many existing machine learning models fall short in streaming contexts due to their reliance on static datasets and sensitivity to contamination. This thesis proposes a unified framework for robust multiple change point detection in streaming environments. The framework integrates Least Squares Support Vector Regression (LS-SVR) with Least Squares Support Vector Data Description (LS-SVDD) to identify deviations in residuals that signal distributional changes. To enhance robustness against outliers and noisy data, the standard LS-SVDD is reformulated using a correntropy-based loss function and optimized via the half-quadratic (HQ) technique. The framework is further extended into an online learning setting, enabling incremental parameter updates and efficient removal of outdated observations without retraining on the full dataset. Extensive simulations and real-world case studies validate the proposed methods, demonstrating superior accuracy, robustness, and computational efficiency compared to existing techniques—providing a practical solution for real-time detection and localization of multiple change points in streaming data. |
| URL: | https://stars.library.ucf.edu/etd2024/487 |
| Database: | OpenDissertations |
| Abstract: | The growing prevalence of high-frequency, high-dimensional data streams in domains such as finance, cybersecurity, and industrial monitoring has intensified the demand for real-time multiple change point detection methods. These methods are expected to identify distributional shifts as they occur while also determining the number and locations of change points—even in the presence of noise, outliers, and limited labeled data. Traditional batch-based approaches and many existing machine learning models fall short in streaming contexts due to their reliance on static datasets and sensitivity to contamination. This thesis proposes a unified framework for robust multiple change point detection in streaming environments. The framework integrates Least Squares Support Vector Regression (LS-SVR) with Least Squares Support Vector Data Description (LS-SVDD) to identify deviations in residuals that signal distributional changes. To enhance robustness against outliers and noisy data, the standard LS-SVDD is reformulated using a correntropy-based loss function and optimized via the half-quadratic (HQ) technique. The framework is further extended into an online learning setting, enabling incremental parameter updates and efficient removal of outdated observations without retraining on the full dataset. Extensive simulations and real-world case studies validate the proposed methods, demonstrating superior accuracy, robustness, and computational efficiency compared to existing techniques—providing a practical solution for real-time detection and localization of multiple change points in streaming data. |
|---|
