Using Open Source Tools for Machine Learning

1. Using Open Source Tools for Machine Learning Samuel Taylor Data Scientist

2. Ryan Born

4. ➔ An introduction to ML ➔ Friendly to newcomers ➔ Helpful to experienced folks ➔ Oriented toward application ➔ Respectful of theory THIS TALK IS

5. THIS TALK IS NOT ➔ A substitute for a Ph. D. ➔ The end-all, be-all ➔ A detailed tutorial

6. Agenda UC2: Forecasting energy load Machine learning intro UC0: Credit card applications UC3: Use ML to ﬁnd your next job UC1: Teach a computer ASL

8. Machine learning?

9. urbanrail.net

10. Machine learning Supervised Unsupervised Other stuﬀ (lots)

11. Machine learning Supervised Classiﬁcation Regression Unsupervised Other stuﬀ (lots)

12. Age Net worth Given credit? 12.5 $500K No 50 $250K No 97 $90K No 50 $750K Yes 53 $650K Yes 60 $500K Yes 62 $800K Yes

17. Machine learning Supervised Classiﬁcation Regression Unsupervised Other stuﬀ (lots)

20. New customer: $500K

21. New customer: $500K

22. New customer: $500K $60K

23. Machine learning Supervised Unsupervised Clustering Other stuﬀ (lots)

25. Group 1 Group 2 Group 3

26. Group 1 Group 2 Group 3

27. Machine learning Supervised Unsupervised Other stuﬀ (lots)

28. ML uses data to approximate something we care about ➔ Goal Find f(x)

29. ML uses data to approximate something we care about ➔ Goal Find f(x) ➔ Problem f(x) is unknown (perhaps unknowable)

30. ML uses data to approximate something we care about ➔ Goal Find f(x) ➔ Problem f(x) is unknown (perhaps unknowable) ➔ But we can measure points from f(x) (with noise)

31. ML uses data to approximate something we care about ➔ Goal Find f(x) ➔ Problem f(x) is unknown (perhaps unknowable) ➔ But we can measure points from f(x) (with noise) ➔ Algorithms to ﬁnd a g(x) that approximates f(x)

33. ➔ What’s the problem? ➔ What does the data look like? ➔ What kind of ML problem is this? ➔ Solution ➔ Lessons learned

34. ➔ What’s the problem? Should we (the bank) give this consumer a credit card?

35. ➔ What does the data look like? Age Net worth (K$) Given credit? 12.5 500 No 50 250 No 62 800 Yes 50 750 Yes 53 650 Yes 60 500 Yes

36. ➔ What kind of ML problem is this?

37. ➔ What kind of ML problem is this? Classiﬁcation

38. ➔ Solution e1701, rpart, igraph, nnet, randomForest, caret, kernlab, ...

39. from sklearn.svm import LinearSVC features = [(12.5, 500), (50, 250), (62, 800), (50, 750), (53, 650), (60, 500)] given_credit = [False, False, True, True, True, True] classifier = LinearSVC() classifier.fit(features, given_credit) classifier.predict([[63, 500]]) ➔ Solution

40. ➔ Solution from sklearn.svm import LinearSVC features = [(12.5, 500), (50, 250), (62, 800), (50, 750), (53, 650), (60, 500)] given_credit = [False, False, True, True, True, True] classifier = LinearSVC() classifier.fit(features, given_credit) classifier.predict([[63, 500]])

45. ➔ Solution from sklearn.svm import LinearSVC features = [(12.5, 500), (50, 250), (62, 800), (50, 750), (53, 650), (60, 500)] given_credit = [False, False, True, True, True, True] classifier = LinearSVC() classifier.fit(features, given_credit) classifier.predict([[63, 500]]) ?

46. ➔ Solution from sklearn.svm import LinearSVC features = [(12.5, 500), (50, 250), (62, 800), (50, 750), (53, 650), (60, 500)] given_credit = [False, False, True, True, True, True] classifier = LinearSVC() classifier.fit(features, given_credit) classifier.predict([[63, 500]]) +

47. How accurate is it?

48. Measuring error Measuring error

49. Measuring error Measuring error

50. Measuring errorRandomly selected testing data allows for model evaluation

51. Measuring errorRandomly selected testing data allows for model evaluation sklearn.model_selection.train_test_split

52. Measuring errorTrain a model on the training data, see how it does on the testing data

55. Measuring error Measuring error ➔ Hold out some “testing data” ➔ Compare test data to prediction ➔ Ideally, calculate real cost

56. Actually a warhead Actually not a warhead Predict a warhead 👍 Destruction of humanity Predict not a warhead Destruction of humanity 👍

57. Actually me Actually not me Recognize ﬁngerprint 👍 People criticize my memes for not being funny Reject ﬁngerprint I get a little annoyed 👍

58. Deﬁning real cost is not always possible

59. Input True Predict Diﬀ Sq. diﬀ 0.53 -8.10 -1.51 -6.60 43.50 4.74 -8.47 -9.60 1.13 1.27 5.79 -16.45 -11.62 -4.83 23.30 9.47 -21.01 -18.70 -2.31 5.34 (43.5 + 1.27 + 23.3 + 5.34) / 4 = 18.35 Mean squared error

60. Input True Predict Error? 0.53 1 1 0 3.68 1 1 0 5.26 0 1 1 7.89 0 0 0 Classiﬁcation error

61. ➔ This stuﬀ is pretty neat ➔ Testing data enables evaluation UC0: LESSONS LEARNED

63. ➔ What’s the problem? I don’t know sign language.

64. khawkins04

66. UC1: Teach a computer ASL joint1_x joint1_y joint1_z ... joint20_x joint20_y joint20_z sign -14.24845886 -11.23913574 47.79299927 ... 39.12654877 -20.38291168 -67.37110138 a -14.24845886 -11.23913574 47.79299927 ... 39.12654877 -20.38291168 -67.37110138 a -14.24845886 -11.23913574 47.79299927 ... 39.12654877 -20.38291168 -67.37110138 a -14.66805267 -12.86016846 47.25432587 ... 39.19580078 -18.27232361 -68.12595367 a -6.099303246 3.211929321 -21.70319366 ... 1.87420845 11.96398926 -98.45552063 b -5.093156815 2.45741272 -22.05827522 ... 6.529464722 14.67698669 -97.91105652 b 32.73310089 -1.139434814 -12.70455551 ... 8.51625061 18.76667786 -97.07907867 b 33.09098053 1.941070557 -11.63526344 ... 10.23889732 31.46665955 -93.68971252 b -23.29023552 -0.6312103271 -21.13870239 ... 14.70001984 23.49594116 -95.80595398 b 32.82236862 -1.860855103 -12.38504791 ... 10.76865768 19.6521759 -96.92489624 b ➔ What does the data look like?

69. ➔ Solution Choose a model ➔ Split data into training, testing ➔ Train a bunch of models on training data ➔ Evaluate them on test data ➔ Select the best one

70. ➔ Solution Build an application ➔ Keyboard… not so great ➔ But! It’s good enough to make an educational game

72. ssaamm/sign-language-tutor

73. ➔ Deﬁne the problem ➔ Limit scope ➔ Model selection ➔ More than the model UC1: LESSONS LEARNED

75. ➔ What’s the problem? Must know when to schedule energy production

76. UC1: Teach a computer ASL

77. UC1: Teach a computer ASL ➔ What does the data look like?

80. ➔ What kind of ML problem is this? Regression

81. Measuring errorTime series data exhibits seasonality

82. Measuring errorFor time series data, random train/test splits leak information

83. Measuring errorFor time series data, random train/test splits leak information

84. Measuring error When using time series data, split based on time

85. Measuring errorSome models learn poorly from time series data

86. https://facebook.github.io/prophet/

90. ➔ Time series data is special ➔ Seasonality ➔ Train/test split – Don’t use random! UC2: LESSONS LEARNED

92. ➔ What’s the problem? Passive job hunting

97. (Data Scientist, sounds_cool=True) (5, 1) ???

98. Engi- neer web Applica- tions sr jr analytics software data developer Sr. Web Applications Developer - Data Analytics 0 1 1 1 0 1 0 1 1 Jr. Software Developer 0 0 0 0 1 0 1 0 1 Sr. Data Engineer 1 0 0 1 0 0 0 1 0 Data data data data 0 0 0 0 0 0 0 4 0

99. (Data Scientist, sounds_cool=True) (1, 0, 0, 1, 0, 0, 0, 0, 0, 1)

100. X = rated_jobs['title'].as_matrix() y = rated_jobs['sounds_cool'].as_matrix() vect = CountVectorizer() Xp = vect.fit_transform(X).toarray() clf = LogisticRegression().fit(Xp, y) new_job_ratings = clf.predict(new_jobs) # array([ 0., 0., 0., 1., 0., 0., 0., 1.])

106. Measuring error Accuracy Classiﬁcation error: 0.197 (awesome!)

107. Measuring error Accuracy Classiﬁcation error: 0.197 (awesome!) But wait…

108. Measuring error Accuracy Classiﬁcation error: 0.197 (awesome!) But wait… Base rate == 0.197

110. Actual 0 Actual 1 Total Predicted 0 400 100 500 Predicted 1 0 0 0 Total 400 100

111. Measuring error Imbalance Better error metrics

112. oversampling undersampling

113. Measuring error End result

114. ➔ Understand the base rate ➔ Simple doesn’t mean ineﬀective ➔ Approximation-generalization tradeoﬀ UC3: LESSONS LEARNED

117. ➔ Understand the base rate ➔ Simple doesn’t mean ineﬀective ➔ Approximation-generalization tradeoﬀ UC3: LESSONS LEARNED also, it’s easier

118. Deep breath, everyone

119. Takeaways ➔ Supervised learning Using past examples to predict a continuous or discrete value

120. Takeaways ➔ Supervised learning Using past examples to predict a continuous or discrete value ➔ Measuring performance Split data into training and testing subsets

121. Takeaways ➔ Supervised learning Using past examples to predict a continuous or discrete value ➔ Measuring performance Split data into training and testing subsets ➔ K.I.S.S. Try the simplest thing that could possibly work

122. Takeaways ➔ Supervised learning Using past examples to predict a continuous or discrete value ➔ Measuring performance Split data into training and testing subsets ➔ K.I.S.S. Try the simplest thing that could possibly work ➔ Test and iterate

124. slides: go.indeed.com/SamuelATO twitter: @SamuelDataT email: [email protected]

Using Open Source Tools for Machine Learning

More Related Content

Similar to Using Open Source Tools for Machine Learning (20)

More from All Things Open (20)

Recently uploaded (20)

Using Open Source Tools for Machine Learning