1. Discovering the Intrinsic Cardinality and Dimensionality of Time Series using MDL BING HU THANAWIN RAKTHANMANON YUAN HAO SCOTT EVANS1 STEFANO LONARDI EAMONN KEOGH DEPARTMENT OF COMPUTER SCIENCE & ENGINEERING REPORTED BY WANG YAWEN

2. Outline  Introduction  Definitions and Notation  MDL Modeling of Time Series  Algorithm  Experimental Evaluation  Complexity  Conclusion

3. Introduction  Choose the best representation and abstraction level  Discover the natural intrinsic representation model, dimensionality and alphabet cardinality of a time series  Select the best parameters for particular algorithms  An important sub-routine in algorithms for classification, clustering and outlier discovery  Minimal Description Length(MDL) fame work

4. Introduction  Dimension reduction  Discrete Fourier Transform(DFT)  Discrete Wavelet Transform(DWT)  Adaptive Piecewise Constant Approximation(APCA)  Piecewise Linear Approximation(PLA)  Choose the best abstraction level and/or representation of the data for a given task/dataset  Useful in its own right to understand/describe the data and an important sub-routine in algorithms for classification, clustering and outlier discovery

5. Introduction  Actual cardinality: 14, 500, 62  Intrinsic cardinality: 2, 2, 12

6. Introduction  Objective  Not simply save memory  Increasing interest in using specialized hardware for data mining, but the complexity of implementing data mining algorithms in hardware typically grows super linearly with the cardinality of the alphabet  Some data mining benefit from having the data represented in the lowest meaningful cardinality

7. Introduction  Objective  Most time series indexing algorithms critically depend on the ability to reduce the dimensionality or the cardinality of the time series, and searching over the compacted representation in main memory  Remove the spurious precision induced by a cardinality/dimensionally that is too high in resource- limited devices  Create very simple outlier detection models

8. Introduction  MDL framework  Automatically discover the parameters that reflect the intrinsic model/cardinality/dimensionally of the data  Without requiring external information or expensive cross validation search

9. Definitions and Notations  MDL is defined for discrete values  Reduce the original number of possible values to a manageable amount  The quantization makes no perceptible difference

10. Definitions and Notations

11.  How many bits it takes to represent a time series T

12. Definitions and Notations  Convert a given time series to other representation or model  DFT, APCA, PLA

13. Definitions and Notations  DL(H): model cost  DL(T|H): correction cost(description cost or error term)  DL(T|H) = DL(T-H)

14. MDL Modeling of Time Series

15.  APCA  Mean 8  16 possible values, DL(H) = 4

16. MDL Modeling of Time Series

18. Algorithm  Discover the intrinsic cardinality and dimensionality of an input time series  Find the right model or data representation for the given time series

19. Algorithm

20.  APCA  Constant lines  Dimensionality: m/2  d constant segments  d-1 pointers to Indicate the offset of the end of each segment

21. Algorithm  PLA  Starting value  Ending value  Ending offset

22. Algorithm  DFT  Linear combination of sine waves  Half set of all coefficients  Subsets of half coef to approximately regenerate T  Sort by absolute value  Use top-d coefficients  inverseDFT  Constant bits(32 bits) for max and min value of the real parts and of the imaginary parts Hence

23. Experimental Evaluation  A detailed example on a famous problem  Baseline  L-Method: explain the residual error vs. size-of-model curve using all possible pairs of two regression lines  10  Bayesian Information Criterion based method  4

24. Experimental Evaluation  An example application in physiology

25. Experimental Evaluation  An example application in astronomy  Anomaly detector

26. Experimental Evaluation  An example application in cardiology

27. Experimental Evaluation  An example application in geosciences

28. Complexity  Space complexity  Linear in the size of the original data  Time complexity  O(mlog 2 m)

29. Conclusion  Simple methodology based on MDL  Robustly specify the intrinsic model, cardinality and dimensionality of time series data from a wide variety of domains  General and parameter-free