The Taub Faculty of Computer Science Events and Talks

Time-series forecasting is widely employed in a variety of domains to predict future trends, tendencies, and properties of the data. However, predicting simple data items is often not enough. Many applications are characterized by a requirement to simultaneously monitor hundreds or even thousands of data series and could benefit from recognizing future occurrences of composite patterns in advance. Despite the rising need for such functionality, this problem received limited attention in recent years. In this work, we formally define and study the problem of predicting patterns over basic data items in multivariate time-series data. Our proposed solution utilizes a combination of a deep learning time-series forecasting model and a complex event processing (CEP) evaluation tree. We also apply attention mechanisms to improve the performance of the forecasting models. We devise a system capable of forecasting composite patterns in a multivariate time-series using a variety of models and suitable for multiple types of data. Our extensive experimental evaluation on three real-world datasets demonstrates the effectiveness and accuracy of our approach.