Time Series Shapelets: A New Primitive for Data Mining

1. Time Series Shapelets: A New Primitive for Data Mining Lexiang Ye and Eamonn Keogh University of California, Riverside

2. Classification • Classification • Huge interest in time series • Extensive applications • Nearest Neighbor • Most accurate (in extensive empirical tests) • Robust • Simple

3. Drawback of the NN • Time and space complexity • Results are not interpretable

4. Solution • Shapelets • shapelets are time series subsequences which are maximally representative of a class • Distinguishing substring selection • Probe design (computational biology)

5. Motivating example

6. false nettles stinging nettles false nettles Shapelet Dictionary I Shapelet 5.1 Leaf Decision Tree I yes no 0 1 false nettles stinging nettles stinging nettles false nettles

7. Brute-Force Algorithm

8. Extract subsequences of all possible lengths Candidates Pool ca . . .

9. Testing the utility of a candidate shapelet • Information gain • Arrange the time series objects • Find the optimal split point • Pick the candidate achieving best utility as the shapelet candidate Split Point 0

10. Candidates Pool Problem • Total number of candidate • Trace dataset • 200 instances, each of length 275 • 7,480,200 shapelet candidates • approximately three days . . .

11. Speedup • Distance calculations from time series objects to shapelet candidates are the most expensive part • Reduce the time in two ways • Distance Early Abandon (known idea) • Admissible Entropy Pruning (novel idea)

12. Admissible Entropy Pruning

13. Admissible Entropy Pruning • Information Gain • Traditional evaluation in decision tree • Easily generalized to the multi-class problem • Reduce the number of distance calculations

14. stinging nettles false nettles 0

15. I=0.42 I= 0.29 0 0

16. false nettles stinging nettles false nettles Shapelet Dictionary I Shapelet 5.1 Classification Leaf Decision Tree I yes no 0 1 false nettles stinging nettles stinging nettles false nettles

17. EXPERIMENTAL EVALUATION

18. Performance Comparison 5 *105 1.00 Brute Force 4 *105 0.95 3 *105 seconds accuracy 0.90 2 *105 Currently best published accuracy 91.1% Pruning 0.85 1 *105 0 0.80 160 10 20 40 80 10 20 40 80 320 160 |D|, the number of objects in the database |D|, the number of objects in the database

19. Projectile Points

20. Arrowhead Decision Tree I II 0 2 1 Avonlea Clovis 1.0 (Clovis) 11.24 I 0 (Avonlea) 85.47 II Shapelet Dictionary 0 200 400

21. Wheat Spectrography 1 0.5 0 0 200 400 600 800 1000 1200 one sample from each class

22. I V II III IV VI 2 4 0 1 3 6 5 Shapelet Dictionary I 0.4 II 0.3 III 0.2 IV 0.1 0.0 V VI 300 0 100 200 Wheat Decision Tree

23. the Gun/NoGun Problem No Gun Gun (No Gun) 2 38.94 I 0 Shapelet Dictionary 0 50 100 Gun Decision Tree I 1 0

24. Gait Analysis

25. 0 100 200 300 Reduces the sensitivity of alignment 1.0 0 0.909 0.902 0.860 right toe 144.075 I left toe (Normal Walk) Walk Decision Tree I 0.535 0 1

26. Conclusions • Interpretable results • more accurate/robust • significantly faster at classification

27. Thank You  Question? • All of the datasets are free to download http://www.cs.ucr.edu/~lexiangy/shapelet.html • Code available upon request