ANOMALY DETECTION IN UNIVARIATE TIME SERIES USING A MULTI-CRITERIA APPROACH
Maciej Wolny · Scientific Papers of Silesian University of Technology Organization and Management Series · 2024
Purpose: The purpose of this study is to propose and evaluate a multi-criteria framework for anomaly detection in univariate time series.By integrating various statistical and machine learning techniques, the study aims to enhance the accuracy and robustness of anomaly identification.This approach seeks to address the challenges posed by complex and dynamic datasets, providing a flexible methodology suitable for diverse applications.Design/methodology/approach: The study employs a multi-criteria approach to anomaly detection in univariate time series.It integrates statistical methods, such as boxplots and deviation-based rules, with machine learning techniques like clustering (hierarchical and k-means).The framework includes three aggregation strategies: restrictive, liberal, and scoring-based, to evaluate anomalies based on different criteria.The methodology is demonstrated using synthetic time series data that incorporates trends, seasonality, noise, and controlled anomalies.Findings: The study reveals significant differences in the performance of various anomaly detection methods applied to univariate time series.Restrictive approaches provide high specificity, minimizing false positives, while liberal methods are more inclusive but prone to false alarms.The scoring-based approach offers a balanced evaluation, enabling quantification of anomaly significance across multiple criteria.The results demonstrate that combining statistical and machine learning methods enhances detection precision.The proposed multicriteria framework is adaptable to diverse applications, though further validation on real-world datasets is required to confirm its effectiveness and scalability.Originality/value: This study introduces a novel multi-criteria framework for anomaly detection in univariate time series, combining statistical and machine learning techniques with aggregation strategies to enhance detection accuracy.Unlike existing approaches, it systematically integrates multiple criteria and evaluates their collective impact on anomaly identification.The framework provides flexibility through restrictive, liberal, and scoring-based methods.Its originality lies in the methodological synthesis and the potential to address complex challenges in anomaly detection across various domains, offering valuable insights for both researchers and practitioners.