Dynamic weight-based granular representation of time series and its application in collective anomaly detection.

In addressing the complexities of time series analysis, two primary challenges emerge: high dimensionality and inherent non-linearity, which often obstruct effective data processing and analysis. Representation learning emerges as a pivotal solution, enabling the efficient transformation of complex, high-dimensional time series into formats conducive to deeper analysis and understanding. This paper proposes a dynamic weight-based granular representation method for time series and validates it in collective anomaly detection tasks. The proposed method innovatively incorporates a weight allocation mechanism within the framework of justifiable granularity, a principle that accentuates the central tendency and extreme features of the data. By representing the original time series through this enhanced principle, the proposed method generates three distinct sets of interval granules: central, maximum value, and baseline granules, each reflecting crucial characteristics of the original data. This granular combination provides a holistic representation of time series, facilitating a more comprehensive analysis and interpretation. Subsequently, the proposed method evaluates the similarity across all subsequences, pinpointing those with notably low similarity as potential anomalies. Through extensive validation, the proposed method has shown high effectiveness in capturing essential time series features such as central tendency and amplitude variations, outperforming existing methods in key metrics including Accuracy, Sensitivity, Specificity, and F1 Score. Furthermore, the Friedman test confirms the proposed method's significant advantages in three of these indicators, showcasing its capability to address the complexities of time series analysis effectively. [ABSTRACT FROM AUTHOR]