Traditional time series analysis focuses on the problem of forecasting future values for a single or multidimensional time series. However, there are a wide range of alternative time series problems that are addressed by time series data mining, including classification, clustering and anomaly detection. These problems arise in all areas of science and engineering. This project will focus on time series classification (TSC). In TSC, the data consists of a set of time series, each with a class label. The problem is to learn a model from the data that can predict the class label for new data. Suppose, for example, we measure a set of patient's blood pressure over time and wish to diagnose an illness based on these measurements alone, or we want to estimate whether any commodity price will rise or fall based on the recent observed price history or volume of trade.  The key feature of TSC is choosing the representation of the data that best captures the difference between classes. We have developed approaches to the TSC problem based on combining shapelets [1,2,3] and other transforms [4] and applied them to a range of problem domains [5]. The project will look at extending this work to classifying multivariate time series. With multivariate time series classification we have multiple time series associated with each class label. So, for example, imagine that in addition to blood pressure we had a collection of observations associated with each patient, such as electrical activity in the brain, heart rate and motion data. Or that rather than use the price movement of a single commodity, we wanted to use the time series from several different economic indicators and other market prices to predict share movements. In either case, we wish to be able to utilise the extra information in the data to construct better classifiers.

The most common data mining approach to dealing with multidimensional time series classification is to simply concatenate the series and use univariate transformations and algorithms. This essentially discards any information about the correlation between series. We will look at specific time series transforms that attempt to capture the relationship between series to construct better classifiers. We will apply the techniques to data generated by our collaborators in areas such as electricity usage profiles, motion capture data, environmental spacial-temporal data and automated lip reading.

L. Ye and E. Keogh, Time series shapelets: a new primitive for data mining, SIGKDD 2009

A. Mueen, E. Keogh and N. Young. Logical-Shapelets: An Expressive Primitive for Time Series Classification, SIGKDD 2012

J. Lines and L. Davis and J. Hills  and A. Bagnall, A Shapelet Transform for Time Series Classification, SIGKDD 2012

A. Bagnall and L. Davis and J. Hills  and J. Lines, Transformation Based Ensembles for Time Series Classification, SDM 2012

L. Davis and B. J. Theobald and A. Toms and A. Bagnall, On the Extraction and Classification of Hand Outlines, International Journal of Neural Systems, 22(4), 2012.