Online Meta-Recommendation of CUSUM Hyperparameters for Enhanced Drift Detection

Area of Science:

• Data Science
• Machine Learning
• Statistical Analysis

Background:

• Time-series analysis is crucial for IoT and real-time systems, requiring accurate change point detection for short-term prediction.
• The Cumulative Sum (CUSUM) method is effective for change detection due to its simplicity and robustness, but its performance relies heavily on hyperparameter tuning.
• Traditional hyperparameter optimization methods for CUSUM are inefficient and subjective, often involving trial-and-error or expert knowledge.

Purpose of the Study:

• To develop an automated meta-modeling scheme for recommending hyperparameters for the CUSUM algorithm.
• To address the limitations of traditional, manual hyperparameter optimization in time-series analysis.
• To improve the efficiency and objectivity of CUSUM hyperparameter selection for dynamic scenarios.

Main Methods:

• Implementation of a meta-modeling framework to automate CUSUM hyperparameter recommendations.
• Evaluation of the proposed framework using benchmark time-series datasets from existing literature.
• Comparison of the meta-modeling approach against traditional optimization techniques like Grid Search and Genetic Algorithms.

Main Results:

• The proposed meta-modeling scheme successfully automates hyperparameter selection for the CUSUM algorithm.
• The framework demonstrates the ability to maintain high accuracy in change point detection.
• Significant reductions in computation time were observed compared to Grid Search and Genetic Algorithm optimization methods.

Conclusions:

• Meta-learning-based techniques offer a viable and efficient alternative for periodic hyperparameter optimization in time-series analysis.
• The developed meta-modeling scheme provides an automated and effective solution for CUSUM hyperparameter tuning.
• This approach enhances the practicality and performance of CUSUM for real-world applications with dynamic data patterns.