The Collective of Transform Ensembles (COTE) for Time Series Classification

Abstracts are not currently available in GtR for all funded research. This is normally because the abstract was not required at the time of proposal submission, but may be because it included sensitive information such as personal details.

Time series classification (TSC) problems involve training a classifier on a set of cases, where each case contains an ordered set of real valued attributes and a class label. The aims of this project are to develop a range of algorithms for TSC that are significantly more accurate and informative than current techniques and to apply these methods to three problem domains of great scientific interest. Our classification technique is based on combining classifiers over alternative data representations. We have shown that ensembling classifiers in the time domain leads to significantly more accurate classifiers than all of the alternative algorithms proposed in the data mining literature. The next logical step is to combine ensembles over different data representations. Our preliminary results found by forming a collective of transform ensembles (COTE), where ensembles are formed in time, frequency, autocorrelation and shapelet space and then combined in a hierarchical ensemble, are extremely promising. Our proposal seeks to build on these preliminary results, find new applications for our algorithm and to extend its impact. One of the driving principles of our research is to implement the algorithms we develop in such a way that it is easy for others to use the code and to reduce the time it takes to perform an exploratory analysis in new problem domains.

unavailable