More information can be found at

All our code is archived on bitbucket.

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. TSC problems arise in a wide range of fields including, but not limited to, data mining, statistics, machine learning, signal processing, environmental sciences, computational biology, image processing and chemometrics.

The MLS group have a strong track record in TSC and have published these papers in this area over the last seven years. We have contributed data sets and results to the UCR time series repository and are continuing to develop new algorithms for a wide range of TSC problems. Our core thesis regarding time series classification is that the best approach is to separate the data transformation from the classification stage.

Similarity in Change

Similarity in change refers to the situation where the relevant discriminatory features are related to the autocorrelation function of each series. The most common approach in this situation is to fit an ARMA model, then base similarity on differences in model parameters. Our contribution to date in this area is has been to focus on speed. In [12,13,14] we demonstrate that by clipping the data (converting a real valued series into binary series of above and below the average) we can approximate the ACF and thus asymptotically perform as well as using the full series. In [2] we extend this idea to show that the run length histogram also approximates the ACF and yet can be calculated in linear time and updated in constant time. We show that classifiers constructed on the run length histogram do not perform significantly worse than those on the ACF. Runlengths

Similarity in Shape

Similarity in shape describes the scenario where class membership is characterised by a common shape but the discriminatory shape is phase independent. If the common shape involves the whole series, but is phase shifted between instances of the same class, then transformation into the frequency domain is one valid approach [15]. If the common shape is local and embedded in confounding noise, then a subsequence techniques such as Shapelets can be employed [1,8,9]. Our contribution was to show that by treating the shapelet algorithm as a transformation rather than as a sub routine for a decision tree, we can significantly increase the classification accuracy. 

Similarity in Time

Similarity in time is characterised by the situation where the series from each class are observations of an underlying common curve in the time dimension. Variation around this underlying common shape is caused by noise in observation, and also by possible noise in indexing which may cause a slight phase shift. Similarity in time can be quantified by measures such as Euclidean distance or correlation, but if there is localised phase shift an elastic measure such as dynamic time warping or edit distance is more appropriate. Our work in this area focuses on evaluating and combining alternative elastic distance measures.


The correct similarity measure/transformation is clearly problem dependent. In [7] we demonstrated that by ensembling transformations we can significantly improve accuracy. We have recently developed an ensemble of elastic distance measures that is significantly more accurate on 75 data sets commonly used in the literature than any other time series classifier yet proposed in the literature. For more details visit Elastic-ensembles.

We have worked with data from a wide range of problem domains, including electricity usage profiles, and bone outline classification and otolith outline classification.

All Projects


  1. Hills, J., Lines, J, Baranauskas, E., Mapp, J. and Bagnall, A. Classification of time series by shapelet transformation. Data Mining and Knowledge Discovery Journal. ISSN 1384-5810, 2013.
  2. Bagnall, A. and Janacek, G., A Run Length Transformation for Discriminating Between Auto Regressive Time Series. Journal of Classification. ISSN 0176-4268, 2013.
  3. Mapp, J, Fisher, M, Bagnall, T, Lines, J, , Songer, S and Scutt Phillips, J. Clupea Harengus: Intraspecies Distinction Using Curvature Scale Space and Shapelets. In: Proc. 2nd International Conference on Pattern Recognition Applications and Methods, ICPRAM 2013.
  4. Hills, J, Bagnall, A, De La Iglesia, B. and Richards, G BruteSuppression: a size reduction method for Apriori rule sets. Journal of Intelligent Information Systems. ISSN 0925-9902, 2013.
  5. Davis, L, Theobald, BJ and Bagnall, A. Automated Bone Age Assessment Using Feature Extraction. Lecture Notes in Computer Sciences, 7435. pp. 43-51. ISSN 0302-9743, 2012.
  6. Davis, LM, Theobald, BJ, Lines, J, Toms, A and Bagnall, A. On the segmentation and classification of hand radiographs. International Journal of Neural Systems, 22 (05). pp. 1250020-1250036. ISSN 0129-0657, 2012.
  7. Bagnall, A, Davis, L, Hills, J and Lines, J., Transformation Based Ensembles for Time Series Classification. In: Proceedings of the SIAM International Conference on Data Mining, 2012-01-01, Anaheim, California, 2012.
  8. Lines, J, Davis, L, Hills, J and Bagnall, A., A Shapelet Transform for Time Series Classification. In: Proceedings of the 18th International Conference on Knowledge Discovery in Data and Data Mining, 2012.
  9. Lines, J and Bagnall, A. Alternative Quality Measures for Time Series Shapelets. Lecture Notes in Computer Sciences, 7435. pp. 475-483. ISSN 0302-9743, 2012.
  10. Lines, JA, Bagnall, AJ, Caiger-Smith, P and Anderson, S Classification of Household Devices by Electricity Usage Profiles. Proceedings of the 12th International Conference on Intelligent Data Engineering and Automated Learning. pp. 403-412, 2011.
  11. Bull, L, Studley, M, Bagnall, AJ and Whittley, IM (2007) Learning Classifier System Ensembles With Rule-Sharing. IEEE Transactions on Evolutionary Computation, 11 (4). pp. 496-502. ISSN 1089-778X, 2007.
  12. Bagnall, AJ, Ratanamahatan, C, Keogh, E, Lonardi, S and Janacek, GJ (2006) A Bit Level Representation for Time Series Data Mining with Shape Based Similarity. Data Mining and Knowledge Discovery Journal, 13 (1). pp. 11-40. ISSN 1384-5810, 2006.
  13. Bagnall, A. J. and Janacek, G. J. Clustering time series with clipped data. Machine Learning, 58 (2). pp. 151-178. ISSN 0885-6125, 2005.
  14. Bagnall, AJ and Janacek, G.J. Clustering time series from ARMA models with clipped data. In: KDD '04 Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining, 2004.
  15. Bagnall, A. and Janacek, G. A likelihood ratio distance measure for the similarity between the Fourier transform of time series. Advances in Knowledge Discovery and Data Mining, 737-743

Research Team

Dr. Tony Bagnall, Dr. Gareth Janacek, Luke Davis, Jason Lines, Jon Hills