Time series describe dynamic processes. Driven by big data applications including mapping of land use from satellite observations over time, our award winning research is revolutionising time series classification by developing technologies that can learn from and accurately classify orders of magnitude larger time series collections than the previous state of the art.
ROCKET and its successors MiniROCKET, MultiROCKET and HYDRA use convolutional filters from deep learning to extract diverse time series features of types that have previously each been addressed by specialised techniques. ROCKET generates a many of these filters and uses them to extract features from each series. From these features a simple linear classifier can learn models that are as accurate as the prior state-of-the-art, but do so in a fraction of the time and create models that classify with blistering speed. Angus Dempster received The Computing Research and Education Association of Australasia Distinguished Dissertation Award for this research. An implementation can be downloaded here. The most recent paper can be found here. Angus' video explaining ROCKET and its successors can be found here.
QUANT is a highly efficient interval method time series classifier, assessed by a recent benchmarking paper as best interval classifier and as "achieving high accuracy remarkably fast." An implementation can be downloaded here. The paper can be found here.
Proximity Forest provides a significant advance on the state of the art in time series classification. By coupling the efficiency of divide and conquer tree classifiers with the effectiveness of specialised similarity measures specifically designed for time series, Proximity Forest achieves very high accuracy for modest computation. An implementation can be downloaded here. The most recent paper can be found here.
TS-Chief builds upon Proximity Forest, enhancing its proximity-based methods by integrating interval statistics and dictionary techniques. An implementation can be found here and the paper found here.
InceptionTime brings the power of deep learning to time series classification. An implementation can be downloaded here. The paper can be downloaded here.
LB Webb and LB Enhanced are our novel lower bounds for Dynamic Time Warping that are both faster and tighter than the popular LB_Keogh. Implementations can be downloaded here and here. The papers can be found here and here.
The following is a blog post on the use of Barycentric averaging in time series classification: http://www.kdnuggets.com/2014/12/averaging-improves-accuracy-speed-time-series-classification.html. The code can be downloaded here: http://francois-petitjean.com/Research/ICDM2014-DTW/index.php. The slides for the 10-year Highest Impact Paper Award winning ICDM 2014 paper can be downloaded here: http://francois-petitjean.com/Research/ICDM2014-DTW/Slides.pdf.
The TSI software for the SDM 2017 paper on time series indexing can be downloaded here: https://github.com/ChangWeiTan/TSI. Slides for the SDM 2017 paper can be found here: http://francois-petitjean.com/Research/SDM17-slides.pdf.
The software for the Best Paper Award winning SDM 2018 paper on finding the best warping window can be downloaded here: https://github.com/ChangWeiTan/FastWWSearch (Matlab version).
Resources referred to in Deep Learning for Time Series Classification and Extrinsic Regression: A Current Survey can be found here: https://github.com/Navidfoumani/TSC_Survey.
Publications
[bibtex file=publications.bib template=publist-time-series format=my-format sort=year order=desc process_titles=0]