Modern Time Series Analysis with STUMPY

Traditional time series analysis techniques have found success in a variety of data mining tasks. However, they often require years of experience to master and the recent development of straightforward, easy-to-use analysis tools has been lacking. STUMPY is a scientific Python library for modern time series analysis that leverages popular open source software and enables you to do better science!

Sean Law is a senior applied scientific researcher and data scientist currently working with a multi-talented Exploration Lab team and serves as an advisor on an enterprise A.I. Council at TD Ameritrade. He has experience producing cutting edge methodologies, building high-performance predictive models, and developing rapid prototypes. Additionally, he is one of the co-organizers of PyData Ann Arbor and is also the creator and core maintainer of STUMPY, a powerful and scalable open source Python library that can be used for a variety of time series data mining tasks.

Transcript

1. Modern Time Series Analysis With #PyDataGlobal Sean M. Law @seanmylaw @stumpy_dev November 2020
2. #ThankYou
3. EXPLORATION LAB TECHNOLOGY Senior Applied Research Scientist and Lead Data Scientist
4. The Problem @seanmylaw
5. #UhOh Here’s some time series data…
6. Now, perform some magic, Dr. McFancyPants! #UhOh Here’s some time series data…
7. #RawData n = 1,000 n = 100 n = 10
8. #RawData n = 10,000
9. #OverwhelmingDread
10. <Insert Magic Spell>
11. #MagicSpells Dynamic Time Warping Statistics Anomaly Detection Visualizations ARIMA Clustering Machine Learning Forecasting
12. Deep Learning #HammerTime
13. No Silver Bullet #NoFreeLunch
14. #IllustrativeExample 0 1 3 2 9 1 14 15 1 2 2 10 7 Time Series with Length, n = 13
15. subsequence /ˈsəbsəkwəns/ noun a part or section of the full time series
16. #IllustrativeExample 0 1 3 2 9 1 14 15 1 2 2 10 7 Subsequence
17. #IllustrativeExample 0 1 3 2 9 1 14 15 1 2 2 10 7 Subsequence
18. #IllustrativeExample 0 1 3 2 9 1 14 15 1 2 2 10 7 Subsequence
19. What’s the Goal? @seanmylaw
20. If a behavior is conserved, there must have been a reason why it was conserved and teasing out these reasons/causes is often very useful… #Goal
21. Do any conserved behaviors exist in my time series data? If there are conserved behaviors, what are they and where are they?
22. Occam’s Razor “simpler solutions are more likely to be correct than complex ones”
23. #KeyQualities What’s the most simple and intuitive approach? Easy To Interpret User/Data Agnostic Parameter Free No Prior Knowledge
24. #Parameter 0 1 3 2 9 1 14 15 1 2 2 10 7 Choose Subsequence Length, m
25. #IllustrativeExample 0 1 3 2 9 1 14 15 1 2 2 10 7 Compare Subsequences
26. Euclidean distance /yo͞oˈklidēən/ /ˈdistəns/ noun the straight-line distance between two points
27. #Back2School 0 1 3 2 9 1 14 15 1 2 2 10 7 Euclidean Distance
28. #Back2School 0 1 3 2 9 1 14 15 1 2 2 10 7 Euclidean Distance (0,1) (1,2) Euclidean Distance X Y x1 x2 y1 y2
29. #Back2School 0 1 3 2 9 1 14 15 1 2 2 10 7 Euclidean Distance (0,1) (1,2) H X Y x1 x2 y1 y2
30. #Back2School 0 1 3 2 9 1 14 15 1 2 2 10 7 Euclidean Distance (0,1) (1,2) H A B X Y x1 x2 y1 y2
31. #Back2School 0 1 3 2 9 1 14 15 1 2 2 10 7 Pythagorean Theorem (0,1) (1,2) H A B X Y x1 x2 y1 y2
32. #Back2School 0 1 3 2 9 1 14 15 1 2 2 10 7 Pythagorean Theorem (0,1) (1,2) H A B X Y x1 x2 y1 y2
33. #Back2School 0 1 3 2 9 1 14 15 1 2 2 10 7 Pythagorean Theorem (0,1) (1,2) H A B X Y x1 x2 y1 y2
34. #Back2School 0 1 3 2 9 1 14 15 1 2 2 10 7 Pythagorean Theorem (0,1) (1,2) H A B X Y x1 x2 y1 y2
35. #Back2School 0 1 3 2 9 1 14 15 1 2 2 10 7 Euclidean Distance (0,1,3,2) (1,2,2,10) H x1 x2 y1 y2 x3 x4 y3 y4
36. #DistanceProfile 0 1 3 2 9 1 14 15 1 2 2 10 7 Pairwise Euclidean Distance 0 1 3 2 9 1 14 15 1 2 2 10 7
37. #DistanceProfile 0 1 3 2 9 1 14 15 1 2 2 10 7 Pairwise Euclidean Distance 0.0 7.4 6.9 14.7 19.3 17.7 19.9 15.0 8.2 8.9 0 1 3 2 9 1 14 15 1 2 2 10 7 Distance Profile O(nm)
38. #TrivialMatch 0 1 3 2 9 1 14 15 1 2 2 10 7 Pairwise Euclidean Distance 0.0 7.4 6.9 14.7 19.3 17.7 19.9 15.0 8.2 8.9 0 1 3 2 9 1 14 15 1 2 2 10 7
39. #NextBestMatch 0 1 3 2 9 1 14 15 1 2 2 10 7 Pairwise Euclidean Distance 0.0 7.4 6.9 14.7 19.3 17.7 19.9 15.0 8.2 8.9 0 1 3 2 9 1 14 15 1 2 2 10 7
40. #NextBestMatch 0 1 3 2 9 1 14 15 1 2 2 10 7 Pairwise Euclidean Distance 0.0 7.4 6.9 14.7 19.3 17.7 19.9 15.0 8.2 8.9 0 1 3 2 9 1 14 15 1 2 2 10 7
41. #DistanceMatrix * * * * * * * * * 6.9 * * * * * * * * * 1.4 * * * * * * * * * 6.2 * * * * * * * * * 7.9 * * * * * * * * * 11.4 * * * * * * * * * 13.6 * * * * * * * * * 14.1 * * * * * * * * * 14.0 * * * * * * * * * 1.4 * * * * * * * * * 6.2 0 1 3 2 9 1 14 15 1 2 2 10 7 0 1 3 2 9 1 14 15 1 2 2 10 7
42. #DistanceMatrix Stacked Distance Profiles * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * 6.9 1.4 6.2 7.9 11.4 13.6 14.1 14.0 1.4 6.2 Distance Matrix
43. #KeyQualities What’s the most simple and intuitive approach? Easy To Interpret User/Data Agnostic Parameter Free No Prior Knowledge * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * 6.9 1.4 6.2 7.9 11.4 13.6 14.1 14.0 1.4 6.2
44. Easy To Interpret Minimal Assumptions Parameter Free No Prior Knowledge #KeyQuestions What’s the most simple and intuitive approach?
45. #KeyQuestions Is the solution scalable?
46. #DistanceMatrix Brute Force Calculation * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * 6.9 1.4 6.2 7.9 11.4 13.6 14.1 14.0 1.4 6.2
47. #DistanceMatrix Brute Force Calculation * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * 6.9 1.4 6.2 7.9 11.4 13.6 14.1 14.0 1.4 6.2 ts = np.array([ 0., 1., 3., 2., 9., 1., 14., 15., 1., 2., 2., 10., 7.]) n = len(ts) # 13 m = 4 for i in range(n-m+1): for j in range(n-m+1): distance = 0 for k in range(m): distance += (ts[i+k]-ts[j+k])2 distance = math.sqrt(distance) 0 1 3 2 9 1 14 15 1 2 2 10 7 0 1 3 2 9 1 14 15 1 2 2 10 7
48. #DistanceMatrix Brute Force Calculation * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * 6.9 1.4 6.2 7.9 11.4 13.6 14.1 14.0 1.4 6.2 ts = np.array([ 0., 1., 3., 2., 9., 1., 14., 15., 1., 2., 2., 10., 7.]) n = len(ts) # 13 m = 4 for i in range(n-m+1): for j in range(n-m+1): distance = 0 for k in range(m): distance += (ts[i+k]-ts[j+k])2 distance = math.sqrt(distance) 0 1 3 2 9 1 14 15 1 2 2 10 7 0 1 3 2 9 1 14 15 1 2 2 10 7
49. #DistanceMatrix Brute Force Calculation * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * 6.9 1.4 6.2 7.9 11.4 13.6 14.1 14.0 1.4 6.2 ts = np.array([ 0., 1., 3., 2., 9., 1., 14., 15., 1., 2., 2., 10., 7.]) n = len(ts) # 13 m = 4 for i in range(n-m+1): for j in range(n-m+1): distance = 0 for k in range(m): distance += (ts[i+k]-ts[j+k])2 distance = math.sqrt(distance)
50. #DistanceMatrix Brute Force Calculation * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * 6.9 1.4 6.2 7.9 11.4 13.6 14.1 14.0 1.4 6.2 ts = np.array([ 0., 1., 3., 2., 9., 1., 14., 15., 1., 2., 2., 10., 7.]) n = len(ts) # 13 m = 4 for i in range(n-m+1): for j in range(n-m+1): distance = 0 for k in range(m): distance += (ts[i+k]-ts[j+k])2 distance = math.sqrt(distance) O(n2m) O(n2)
51. Redefining #BackOfTheEnvelope n = 5 years x 365 days/year x 24 hours/day x 60 mins/hour x 20 times/min n = 52,560,000 data points 1,598.7 days (4.4 years) and 11.1 PB memory to compute! (n2-n)/2 1,381,276,773,720,000 x 0.000 000 1 seconds Brute Force Calculation 32 bit
52. #KeyQuestions Is the solution scalable? What’s the most simple and intuitive approach?
53. #MagicSpells Dynamic Time Warping Statistics Anomaly Detection Visualizations ARIMA Clustering Machine Learning Forecasting
54. #RedefiningPossible Until Now
55. #LiteratureReview Dec 2016 Today Matrix Profile
56. matrix profile /ˈmātriks/ /ˈprōˌfīl/ noun a vector that stores the distance* between each subsequence within a time series and its nearest neighbor * z-normalized Euclidean distance
57. Distance Matrix Matrix Profile * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * 6.9 1.4 6.2 7.9 11.4 13.6 14.1 14.0 1.4 6.2 6.9 1.4 6.2 7.9 11.4 13.6 14.1 14.0 1.4 6.2
58. Distance Matrix Matrix Profile * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * 6.9 1.4 6.2 7.9 11.4 13.6 14.1 14.0 1.4 6.2 6.9 1.4 6.2 7.9 11.4 13.6 14.1 14.0 1.4 6.2 Space O(n) Time O(n2)
59. Do any conserved behaviors exist in my time series data? If there are conserved behaviors, what are they and where are they?
60. #Motif 0 1 3 2 9 1 14 15 1 2 2 10 7 6.9 1.4 6.2 7.9 11.4 13.6 14.1 14.0 1.4 6.2
61. #Motif 0 1 3 2 9 1 14 15 1 2 2 10 7 6.9 1.4 6.2 7.9 11.4 13.6 14.1 14.0 1.4 6.2 Motif
62. matrix profile index /ˈmātriks/ /ˈprōˌfīl/ /ˈinˌdeks/ noun the index (location) of the nearest neighbor for a given subsequence
63. 6.9 1.4 6.2 7.9 11.4 13.6 14.1 14.0 1.4 6.2 2 8 9 1 9 2 7 2 1 2 0 1 2 3 4 5 6 7 8 9 10 11 12 0 1 3 2 9 1 14 15 1 2 2 10 7 #MatrixProfileIndex
64. 6.9 1.4 6.2 7.9 11.4 13.6 14.1 14.0 1.4 6.2 2 8 9 1 9 2 7 2 1 2 0 1 2 3 4 5 6 7 8 9 10 11 12 0 1 3 2 9 1 14 15 1 2 2 10 7 #MatrixProfileIndex
65. #Discord 0 1 3 2 9 1 14 15 1 2 2 10 7 6.9 1.4 6.2 7.9 11.4 13.6 14.1 14.0 1.4 6.2 Discord
66. #MatrixProfilePerformance STAMP FFT Time O(n2logn) Space O(n) GPU-STOMP Hardware Time O(n2) Space O(n) Brute Force Naive Time O(n2m) Space O(n2) STOMP Algebra Time O(n2) Space O(n)
67. n = 100,000,000 499,999,999,500,000,000 x 0.000 000 1 seconds = 1585.49 years GPU-STOMP 12.13 days
68. Given the matrix profile, most time series data mining problems are trivial or easy to solve in a few lines of code #EamonnKeogh
69. These are the best ideas in time series data mining in the last two decades. #EamonnKeogh
70. Extraordinary claims require extraordinary evidence. #CarlSagan
71. @stumpy_dev
72. STUMPY /ˈstəmpē/ noun a powerful and scalable Python library that efficiently computes the matrix profile, which can be used for a variety of time series data mining tasks
73. ~50K+ Downloads/Installs ~1.5K+ Github Stars #Growth #ThankYou
74. Version 1.5.0 (Nov 2020) conda install -c conda-forge stumpy python -m pip install stumpy @stumpy_dev
75. Minimal Dependencies (3) #Winning Python 3.6+ Core Dask Distribute + + + NumPy Numerical Numba Parallelize + SciPy Accessory
76. State of the Art Days 0 3.25 6.5 9.75 13 Today (Distributed 256 CPUs) GPU-STOMP 12.13 Days 9.75 Days n = 100,000,000
77. #TreeTown STAMP STOMP STUMPED GPU-STUMP
78. #KeyQuestions Is the solution scalable? What’s the most simple and intuitive approach?
79. #TryIt ts = np.random.rand(10_000) m = 50 stumpy.stump(ts, m) stumpy.gpu_stump(ts, m)
80. #TreeTown STAMP STOMP STUMPED GPU-STUMP MSTUMPED SCRUMP STUMPI
81. Why Should You Use STUMPY? @seanmylaw
82. #Why Interpretable Euclidean Distance Complementary Mix-and-Match User Friendly Developed For (Data) Scientists Fast & Scalable Multi-CPU/GPU Support Reliable 100% Code Coverage
83. When Should You Use STUMPY? @seanmylaw
84. #When Dynamic Time Warping Statistics Anomaly Detection Visualizations ARIMA Clustering Machine Learning Forecasting
85. No Silver Bullet #NoFreeLunch
86. Live Demo & Tutorial Highlights @seanmylaw
87. #STUMPY And Moar! Multi- dimensional Semantic Segmentation Motif/Discord Discovery Time Series Chains Shapelets & Snippets MPdist Clustering ML Features
88. Contribute Communicate Consume Final Calls To Action @seanmylaw
89. search: “python stumpy” tutorials: https://tiny.cc/stumpy-tutorials demo: https://tiny.cc/stumpy-demo gpu: https://tiny.cc/stumpy-colab code: https://tiny.cc/stumpy-code docs: https://tiny.cc/stumpy-docs pubs: https://tiny.cc/stumpy-pubs @seanmylaw @stumpy_dev
92. Exclusion Zone @seanmylaw
93. #TrivialMatch Trivial Match Only * 7.4 14.7 19.3 17.7 19.9 15.0 8.2 8.9 7.4 * 10.9 7.9 15.7 18.8 19.1 15.8 8.4 6.9 10.9 * 16.8 16.1 13.6 18.8 14.0 11.6 14.7 16.8 * 16.8 19.8 18.0 19.4 8.2 13.4 19.3 15.7 16.1 16.8 * 20.7 23.6 18.7 15.3 17.7 18.8 19.8 20.7 * 19.2 23.1 19.8 14.4 19.9 19.1 18.8 18.0 23.6 19.2 * 20.1 20.5 15.0 15.8 19.4 18.7 23.1 14.1 * 16.2 16.1 8.2 11.6 8.2 15.3 19.8 20.1 16.2 * 8.6 8.9 8.4 13.4 11.4 14.4 20.5 16.1 8.6 * 6.9 1.4 6.2 7.9 11.4 13.6 14.1 14.0 1.4 6.2 Exclusion Zone * 7.4 14.7 19.3 17.7 19.9 15.0 8.2 8.9 7.4 * 10.9 7.9 15.7 18.8 19.1 15.8 8.4 6.9 10.9 * 16.8 16.1 13.6 18.8 14.0 11.6 14.7 16.8 * 16.8 19.8 18.0 19.4 8.2 13.4 19.3 15.7 16.1 16.8 * 20.7 23.6 18.7 15.3 17.7 18.8 19.8 20.7 * 19.2 23.1 19.8 14.4 19.9 19.1 18.8 23.6 19.2 * 20.1 20.5 15.0 15.8 19.4 18.7 23.1 14.1 * 16.2 16.1 8.2 11.6 8.2 15.3 19.8 20.1 16.2 * 8.6 8.9 8.4 13.4 11.4 14.4 20.5 16.1 8.6 * 6.9 1.4 6.2 7.9 11.4 13.6 14.1 14.0 1.4 6.2 18.0 * 7.4 14.7 19.3 17.7 19.9 15.0 8.9 7.4 * 10.9 7.9 15.7 18.8 19.1 15.8 8.4 6.9 10.9 * 16.8 16.1 13.6 18.8 14.0 11.6 14.7 16.8 * 16.8 19.8 18.0 19.4 13.4 19.3 15.7 16.1 16.8 * 20.7 23.6 18.7 15.3 17.7 18.8 19.8 20.7 * 19.2 23.1 19.8 14.4 19.9 19.1 18.8 23.6 19.2 * 20.1 20.5 15.0 15.8 19.4 18.7 23.1 14.1 * 16.2 16.1 8.2 11.6 8.2 15.3 19.8 20.1 16.2 * 8.6 8.9 8.4 13.4 11.4 14.4 20.5 16.1 8.6 * 6.9 7.9 14.1 8.2 8.2 18.0 1.4 6.2 11.4 13.6 14.0 1.4 6.2 * 7.4 14.7 19.3 17.7 19.9 15.0 8.9 7.4 * 10.9 7.9 15.7 18.8 19.1 15.8 8.4 6.9 10.9 * 16.8 16.1 13.6 18.8 14.0 11.6 14.7 16.8 * 16.8 19.8 18.0 19.4 13.4 19.3 15.7 16.1 16.8 * 20.7 23.6 18.7 15.3 17.7 18.8 19.8 20.7 * 19.2 23.1 19.8 19.9 19.1 18.0 23.6 19.2 * 20.1 20.5 15.0 15.8 19.4 18.7 23.1 14.1 * 16.2 16.1 8.2 11.6 8.2 15.3 19.8 20.1 16.2 * 8.6 8.9 8.4 13.4 11.4 14.4 20.5 16.1 8.6 * 6.9 7.9 13.6 14.1 8.2 8.2 14.4 18.8 1.4 6.2 11.4 14.0 1.4 6.2 * 7.4 14.7 19.3 17.7 19.9 15.0 8.9 7.4 * 10.9 7.9 15.7 18.8 19.1 15.8 8.4 6.9 10.9 * 16.8 16.1 13.6 18.8 14.0 11.6 14.7 16.8 * 16.8 19.8 18.0 19.4 13.4 19.3 15.7 16.1 16.8 * 20.7 23.6 18.7 15.3 17.7 18.8 19.8 20.7 * 19.2 23.1 19.8 14.4 19.9 18.8 18.0 23.6 19.2 * 20.1 20.5 15.0 15.8 19.4 18.7 23.1 14.1 * 16.2 16.1 8.2 11.6 8.2 15.3 19.8 20.1 16.2 * 8.6 8.9 8.4 13.4 11.4 14.4 20.5 16.1 8.6 * 6.9 7.9 13.6 14.1 8.2 8.2 19.1 1.4 6.2 11.4 14.0 1.4 6.2 * 7.4 14.7 19.3 17.7 19.9 15.0 8.9 7.4 * 10.9 7.9 15.7 18.8 19.1 15.8 8.4 6.9 10.9 * 16.8 16.1 13.6 18.8 14.0 11.6 14.7 16.8 * 16.8 19.8 18.0 19.4 8.2 19.3 15.7 16.1 16.8 * 20.7 23.6 18.7 15.3 17.7 18.8 19.8 20.7 * 19.2 23.1 19.8 14.4 19.1 18.8 18.0 23.6 19.2 * 20.1 20.5 15.8 19.4 18.7 23.1 14.1 * 16.2 16.1 8.2 11.6 8.2 15.3 19.8 20.1 16.2 * 8.6 8.9 8.4 13.4 11.4 14.4 20.5 16.1 8.6 * 6.9 7.9 11.4 13.6 14.1 14.0 8.2 13.4 19.9 15.0 1.4 6.2 1.4 6.2 * 7.4 14.7 19.3 17.7 19.9 15.0 8.9 7.4 * 10.9 7.9 15.7 18.8 19.1 15.8 8.4 6.9 10.9 * 16.8 16.1 13.6 18.8 14.0 11.6 14.7 16.8 * 16.8 19.8 18.0 19.4 8.2 13.4 19.3 15.7 16.1 16.8 * 20.7 23.6 18.7 15.3 17.7 18.8 19.8 20.7 * 19.2 23.1 19.8 14.4 19.9 19.1 18.8 18.0 23.6 19.2 * 20.1 20.5 15.8 19.4 18.7 23.1 14.1 * 16.2 16.1 8.2 11.6 8.2 15.3 19.8 20.1 16.2 * 8.6 8.9 8.4 13.4 11.4 14.4 20.5 16.1 8.6 * 6.9 7.9 11.4 13.6 14.1 14.0 8.2 15.0 1.4 6.2 1.4 6.2 i±1 i±2 i±3 i±4 i±5 i±6
94. z-normalize /zē/ /ˈnôrməˌlīz/ verb a vector that stores the z-normalized Euclidean distance between each subsequence and its nearest neighbor
95. #IllustrativeExample 0 1 3 2 9 1 14 15 1 2 2 10 7 Z-normalize -2.875 -3.875 3.125 -4.875 8.125 9.125 -4.875 -3.875 -0.520 -0.700 0.565 -0.881 1.469 1.650 -0.881 -0.700 Subtract the Subsequence Mean (5.875) Divide by the Subsequence Standard Deviation (5.533)
96. Left and Right Matrix Profile Indices @seanmylaw
97. 6.9 1.4 6.2 7.9 11.4 13.6 14.1 14.0 1.4 6.2 2 8 9 1 9 2 7 2 1 2 0 1 2 3 4 5 6 7 8 9 10 11 12 0 1 3 2 9 1 14 15 1 2 2 10 7 -1 0 0 1 1 2 3 2 1 2 Left Matrix Profile Indices
98. 6.9 1.4 6.2 7.9 11.4 13.6 14.1 14.0 1.4 6.2 2 8 9 1 9 2 7 2 1 2 0 1 2 3 4 5 6 7 8 9 10 11 12 0 1 3 2 9 1 14 15 1 2 2 10 7 -1 0 0 1 1 2 3 2 1 2 2 8 9 8 9 9 7 9 9 -1 Left Matrix Profile Indices Right Matrix Profile Indices
99. Matrix Profile Features @seanmylaw
100. What Makes It So Fast?! @seanmylaw
101. #Distance 0 1 3 2 9 1 14 15 1 2 2 10 7 Brief Recap Standard Dot Product!
102. #DotProduct Reusing Dot Products 0 1 3 2 9 1 14 15 1 2 2 10 7 0 1 3 2 9 1 14 15 1 2 2 10 7 * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * 1 0 1 1 3 3 2 2 9 2 + + + 1 3 3 2 2 9 9 1 + + + 3 2 2 9 9 1 1 14 + + + 3 4 5 6 7 8 9 15 1 1 2 2 2 2 10 + + + 1 2 2 2 2 10 10 7 + + +
103. #DotProduct Reusing Dot Products 0 1 3 2 9 1 14 15 1 2 2 10 7 0 1 3 2 9 1 14 15 1 2 2 10 7 * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * 0 1 1 3 3 2 2 9 + + + 1 3 9 1 + + Time O(n2) 0 1 3 2 9 1 14 15 1 2 2 10 7 0 1 3 2 9 1 14 15 1 2 2 10 7 Xi Yj Xi-1 Yj-1 i j
104. #NoTrivialMatch 0 1 3 2 9 1 14 15 1 2 2 10 7 7 4 0 1 3 2 8 * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * Sidebar: AB-Joins 0.0 1.0 6.9 17.7 13.1 8.8 2.2 8.2 16.2 8.6 stumpy.stump(ts_A, m, ts_B) or stumpy.stump(ts_B, m, ts_A)
105. STUMPY Performance @seanmylaw
106. #BiasedPerformance
107. #BiasedPerformance
108. #BiasedPerformance
109. #BiasedPerformance
110. GPU Lessons Learned @seanmylaw
111. #GPU
112. Avoid Additional Dependencies
113. Right Tools, Right Job #Winning Python Numba Dask Core Parallelize Distribute + + + NumPy Numerical GPU Support
114. What’s The Catch?
115. Pros No additional dependencies Code in Python Compiles to CUDA Excellent host-device API Multi-GPU Cons Separate code targeting GPUs GPU multithreading paradigm Delayed CUDA functionality Memory Limitations Writing unit tests and CI is hard!