Universal Event Detection in Time Series
Menouar Azib, B. Renard, Philippe Garnier, Vincent Génot, Nicolás André · 2023
Event detection in time series data is a crucial task spanning various domains, and extensive research has explored methods to achieve this goal. These methods range from traditional threshold-based techniques to more advanced deep learning approaches. However, a comprehensive survey of existing methods reveals that each approach has its limitations, often lacking mathematical validation and exhibiting limited robustness. To address these limitations, this paper introduces a novel framework rooted in universal approximation theory, a well-established and proven methodology. This framework showcases the capability to accurately detect a broad spectrum of events in multivariate time series data with the desired precision. To bolster robustness, the proposed framework employs a stacked ensemble learning meta-model, effectively mitigating individual model weaknesses and biases, thereby yielding more resilient predictions. Moreover, this paper provides a user-friendly quick-start guide for implementing the framework, demonstrated using two diverse datasets: one from the field of planetary science and another from financial security. The results underscore the effectiveness of the proposed framework in achieving high accuracy and precision when detecting events in both datasets. In summary, this paper presents a novel and robust framework for event detection in time series data, founded on universal approximation theory and bolstered by ensemble learning.