CRII: III: Rare Event Prediction in Time Series

点击翻译按钮获取中文摘要

Predicting rare events within a time series is a critical task in many real-world applications. Taking the healthcare industry as an example, it would be very useful to predict an influx in the number of patients that would overwhelm a hospital beyond its capacity. Without such capability, it could have negative consequences in the hospital’s ability to adequate health care. Rare event prediction in time series is a challenging problem due to the non-linear nature of rare events, the inability to capture their key information in the systematic components of the temporal data, and the data imbalance between rare and normal events. This project will result in a framework to address these challenges. With such a prediction framework, steps can be taken to ensure the affected population's readiness for such rare events. For instance, the application of the resulting technology will significantly improve readiness in the healthcare care industry. Furthermore, the resulting technology can be applied to other areas of societal interest.This project studies the nature of rare events in time series and addresses the above-mentioned challenges with three technical aims. First, it is well known that sequence models such as recurrent neural networks can be used to capture both the temporal and non-linear nature of the data. Nevertheless, they have not been well studied for the problem of rare event prediction in time series. One goal of this project is to thoroughly study if sequence models can capture both the temporal and non-linear nature of the rare events to facilitate the prediction. Second, the key information of rare events can be hidden by the seasonal fluctuations in the time series. Another goal of this project is to study if time series decomposition that separates the seasonality can help the prediction of rare events. Finally, in terms of the data imbalance problem, traditional prediction models often treat normal and rare events equivalently, which can be harmful for the rare event prediction. This project will develop a new method based on the properties of rare events to emphasize their importance. Moreover, filters can be specifically designed to extract rare events better to reduce the effect from large amount of normal data. Both strategies have not been well studied for the rare event prediction in time series, which will be the third goal of this project.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.