of Time Series using Minimum Description Length

Supporting Material

- Slide to briefly explain our idea including all figures in higher resolution.

Intrinsic Cardinality and Dimensionality

Donoho-Johnstone Benchmark

- This well-known benchmark is created ten years ago and cited by many publications.
- You can download this benchmark at ftp://ftp.sas.com/pub/neural/dojo/dojo.html.
- We also provide the mirror of the time series in txt file here. Note that this file is required to run the codes for figure 9, 10, and 11.
- Source code for generating the result in figure 9. Matlab figure behind figure 9. (Find an intrinsic dimensionality)
- Source code for generating the result in figure 10. Matlab figure behind figure 10. (Find an intrinsic cardinality)
- Source code for generating the result in figure 11. Matlab figure behind figure 11. (Compare among three models)

Physiology
Dataset

- Above time series is an excerpt data from the Muscle dataset in [1].
- Original data is in the author's webpage; we also provided the mirror of the data here.
- Source code for generating the result in figure 13.
- Matlab figure
to show the result in figure 13.

Ourlier Detection

Astronomy Dataset

- Six of 5,327 star light curves time series in class RRL annotated anomalies by human from [2].
- The mirror of the dataset with labels can be downloaded here.

Cardiology Dataset

- A cardiology data from PhysioBank ATM of the MIT BIH Arrhythmia dataset, record 108, signal MLII.
- The mirror of the timeseries can be downloaded here.

Compare to L-method

- The time series are easily generated by step function with noise. Time series of six steps is shown in figure above.
- We compare our method with L-method; The comparison result shows that our method can find the correct dimensionality in every case.
- We tested on 2-step time series to 16-step time series. These are the results for our method and L-method.

Source Code

- Code for discovering an intrinsic dimensionality of APCA model. (Fix cardinality)
- Code for discovering an intrinsic dimensionality of PLA model. (Fix cardinality)
- Code for discovering an intrinsic dimensionality of DFT model.(Fix cardinality)
- Code for discovering an intrinsic cardinality of APCA model. (Fix dimensionality)

References:

- [1] Mörchen, F. and Ultsch, A. 2005. Optimizing time series discretization for knowledge discovery. KDD, 660-665.
- [2] Protopapas, P., Giammarco, J. M., Faccioli, L., Struble, M. F., Dave, R., and Alcock, C. 2006. Finding outlier light-curves in catalogs of periodic variable stars. Monthly Notices of the Royal Astronomical Society, 369 (2006), 677–696.

Contact Information: rakthant "at" cs "dot" ucr "dot" edu