High performance time series data analysis and pattern discovery

点击翻译按钮获取中文摘要

This research is motivated by the increasing demand for high performance time series data management and analysis techniques. A time series is a sequence of values of a variable measured often at regular time intervals, e.g., yearly, monthly, weekly, daily, and hourly. Advances in technology in development of automatic sensors and devices for data collection and storage have provided us with huge amount of time series that need to be processed to extract the real benefits buried in the data. Time series appear in a wide range of numerous applications in finance, meteorology, telecommunication and wireless networks, security, etc., in which the data is analyzed and searched for similarity matching, correlation queries, and pattern discovery. Two major challenges in management of single and multiple, large scale time series in such applications include modeling and processing, where modeling aims at describing the data and providing insights into its generation mechanisms while processing aims at prediction and forecasting. Time series data has two important characteristics, namely high dimensional and temporal, that affect the performance, explained as follows. Each time point of a time series can be viewed as a dimension, and hence a time series of n values is a point in an n-dimensional space. Processing points in a high dimensional space is extremely difficult, noting that n is large in practice. The temporal property of time series implies that consecutive and near by data values are similar to each other within some predictable threshold. This implies possibility of redundancy and hence opportunity for reduction in dimension and space. This primarily provides storage efficiency, however, its real benefit is time reduction. Storage and time efficient techniques are crucial in success of time series applications. The literature is vast and rich on such techniques for handling Gigabyte or close to Terabyte time series. This research attempts to raise the bar even higher to surpass handling Terabyte data towards Petascale time series. The theoretical and practical developments in this research will advance the state-of-the-art and technology for high performance time series data management.