Identifying the Most Suitable Representation Method for Heterogeneous Time Series Data

摘要

A time series data is a collection of measurements obtained sequentially, which is common in many application domains, e.g.. fluctuations of stock market, observations from sensor networks, medical and biological signals. Since time series data usually contains large number of data points, i.e., high-dimensionality, directly dealing with such data in its raw format is very expensive in terms of processing and storage loading. To effectively and efficiently manage time series data, several representation methods were discussed. Representation methods can reduce the dimensionality of a time series data while preserving its fundamental characteristics. However, each method has its own drawbacks and is most suitable for certain time series data types, which means no single method is efficient enough for all possible types. To address this issue, this study aims at proposing a system that can identify the most suitable representation method for different types of time series. To be specific, this study first proposes a time series clustering approach to cluster sample time series datasets to identify different types of time series. We then conduct an extensive performance evaluation by testing the performance of different representation methods on the clustered time series types. Based on the evaluation, the most suitable representation methods for certain clusters can be identified. With a new time series input, the system can first classify this time series by computing its similarities with clustered time series types, which indirectly helps us identify the representation method that is the most suitable for this new time series data. Finally, evaluation result shows that there are three types of representations are the most suitable representation for different time series types respectively.