[StartupCourse/18] Discover Machine Learning

You want to predict the future ?

Machine Learning is not so complex !

Let's discover it!

Transcript

1. Fabien VAUCHELLES fabien.vauchelles@zelros.com / zelros.com / @fabienv http://bit.ly/scmachinelearning
2. INTRODUCTION
3. I WILL TELL YOU THE TRUTH
4. JACK KNEW HE WOULD DIE
5. YOU HAVE BELIEVED ME ?
6. JACK KNEW HIS FUTURE
7. JACK LIFE DECISION PATH
8. Men Women JACK LIFE DECISION PATH
9. Men Women JACK Child Adult LIFE DECISION PATH
10. Men Women Child Adult 1st class 3th class JACK LIFE DECISION PATH
11. Men Women Child Adult 1st class 3th class GAME OVER JACK LIFE DECISION PATH
12. BUT HE LOVED ROSE...
13. BACK TO MACHINE LEARNING If you know the passengers list: • Gender • Age • Ticket class • Does he survived ? You can create a Decision Tree ... ... for this Supervised Problem !
14. What is Machine Learning
15. With Machine Learning, Computers solve problems which are unsolvable by conventional methods.
16. TYPE OF PROBLEM SUPERVISED LEARNING UNSUPERVISED LEARNING
17. SUPERVISED LEARNING Men Women Child Adult 1st class 3th class GAME OVER JACK
18. UNSUPERVISED LEARNING
19. What is the goal of ML algorithms
20. MINIMIZE THE ERROR
21. THE PROCESS OF DATA SCIENCE BUILD FEATURES VALIDATE THE PREDICTION TRAIN THE MODEL 1 2 3
22. THE PROCESS OF DATA SCIENCE BUILD FEATURES VALIDATE THE PREDICTION TRAIN THE MODEL 1 2 3
23. THE PROCESS OF DATA SCIENCE BUILD FEATURES VALIDATE THE PREDICTION TRAIN THE MODEL 1 2 3
24. BUILD FEATURES
25. What is a feature
26. A feature is a column in a dataset. Age of passengers is a feature !
27. DATA ANALYSIS ANALYZE DATA CLEAN DATA 1 2
28. DATA ANALYSIS ANALYZE DATA CLEAN DATA 1 2
29. DATA ANALYSIS / FREQUENT VALUE Class 2 : 1500 Class 3 : 1200 Class 1 : 1233
30. What is the difference between median and average
31. DATA ANALYSIS / DISTRIBUTION 5% 25% 50% 75% 95% Quantile 1 Quantile 2 Quantile 3 Quantile 4 Outlier BOX PLOT
32. DATA ANALYSIS / DISTRIBUTION
33. DATA ANALYSIS / MISSING VALUES mean(Nan, None, Null, etc) = GAME OVER
34. DATA ANALYSIS / MISSING VALUES Choose replace value carefully ! 10 Nan 20 Nan 30 10 0 20 0 30 Mean = 12 BAD !!!
35. DATA ANALYSIS / MISSING VALUES Fill empty value with median: 20 10 Nan 20 Nan 30 10 20 20 20 30 Mean = 20 GOOD !!!
36. DATA ANALYSIS / REMOVE OUTLIERS Outlier
37. DATA ANALYSIS / REMOVE OUTLIERS Outliers can alter predictions 10 20 30 1000 50 Mean = 222,22 BAD !!!
38. DATA ANALYSIS / REMOVE OUTLIERS Use median: 30 Mean = 28 GOOD !!! 10 20 20 30 30 10 20 30 1000 50
39. MEDIAN IS LESS SENSITIVE TO OUTLIERS THAN AVERAGE
40. TRAIN THE MODEL
41. FIT & PREDICT FIT MODEL PREDICT 1 2
42. FIT & PREDICT FIT MODEL PREDICT 1 2
43. TYPE OF PREDICTION CLASSIFICATION REGRESSION
44. #goatcat #grumpycat #fatcat #killercat #scaredcat #happycat #dumbcat CLASSIFICATION
45. Surface (m2) Rooms Bedrooms Garden (m2) Price (€) 200 5 2 200 500 000 100 3 1 0 200 000 300 5 2 300 800 000 150 4 2 100 300 000 200 4 1 200 ? REGRESSION Find house price:
46. ALGORITHMS DECISION TREE DEEP LEARNING REGRESSION CLUSTERING BAYESIAN NLP Linear Regression Logistic Regression Convolutional Neural Network Deep Boltzmann Machine Recurrent Neural Network Gaussian Naive Bayes Multinomial Naive Bayes Bayesian Network k-Means k-Medians Hierarchical Clustering Perceptron Random Forest Gradient Boosting XGBoost TF-IDF Word2Vec
47. TYPE OF ALGORITHMS REGRESSION DECISION TREES CLUSTERING
49. REGRESSION y X h(X)=θ 0 +θ 1 X / LINEAR
50. REGRESSION / POLYNOMIAL y X h(X)=θ 0 +θ 1 X+θ 2 X2
51. REGRESSION / POLYNOMIAL y X h(X)=θ 0 +θ 1 X+θ 2 X2+θ 3 X3
52. REGRESSION / POLYNOMIAL y X OVERFITTING !!! h(X)=θ 0 +θ 1 X+θ 2 X2+θ 3 X3
53. REGRESSION / COST FUNCTION J=∑(h(X)-y)2 prediction real value Sum of errors : m count
54. REGRESSION / COST FUNCTION h(X)=θ 0 +θ 1 X+θ 2 X2
55. REGRESSION / COST FUNCTION h(X)=θ 0 +θ 1 X+θ 2 X2+θ 3 X3
56. REGRESSION / COST FUNCTION Minimise the error : min(J)
57. REGRESSION / GRADIENT DESCENT
59. DECISION TREES Men Women Child Adult 1st class 3th class GAME OVER JACK / CHOICE
60. DECISION TREES / SPLIT Passenger with class 3 ? Adult passenger ? Male passenger ? Alive Dead
61. DECISION TREES / SPLIT Passenger Class 3 ? YES NO DEAD Passenger with class 3 ? Adult passenger ? Male passenger ? Alive Dead
62. DECISION TREES / SPLIT Adult ? YES NO DEAD Passenger Class 3 ? YES NO DEAD Passenger with class 3 ? Adult passenger ? Male passenger ? Alive Dead
63. DECISION TREES / SPLIT ALIVE DEAD Male ? YES NO Adult ? YES NO DEAD Passenger Class 3 ? YES NO DEAD Passenger with class 3 ? Adult passenger ? Male passenger ? Alive Dead
65. CLUSTERING / GROUP k=2
66. CLUSTERING / DISTANCE Euclidienne Manhattan d(A,B) = |X B -X A |+|Y B -Y A | d(A,B) = sqrt[(X B -X A )2+(Y B -Y A )2]
67. CLUSTERING / CENTROID k=2
68. CLUSTERING / CENTROID k=2
69. CLUSTERING / CENTROID k=2
70. CLUSTERING / CENTROID k=2
71. CLUSTERING / CENTROID k=2
72. CLUSTERING / CENTROID k=2
73. CLUSTERING / CENTROID k=2
74. CLUSTERING / CENTROID k=2
75. CLUSTERING / CENTROID k=2
76. CLUSTERING / CENTROID k=2
77. CLUSTERING / CENTROID k=2
78. VALIDATE THE PREDICTION
79. What should we evaluate errors
80. “I ALWAYS HAVE AN ANSWER FOR YOU”
81. IT’S A GOOD METRIC TO START
82. METRICS FOR REGRESSION J=∑(h(X)-y)2 prediction real value Mean Squared Error : m count
83. METRICS FOR REGRESSION J=∑|h(X)-y| prediction real value Mean Absolute Error : m count
84. METRICS FOR CLASSIFICATION Reality Prediction Jack Dead Alive Rose Alive Alive
85. METRICS FOR CLASSIFICATION Prediction Positive Negative Reality True True positive False negative False False Positive True negative
86. PRECISION Prediction Positive Negative Reality True True positive False negative False False Positive True negative Precision = True positive True positive + False positive
87. “WITH PRECISION, I CAN TELL YOU I WILL DIE, BABY”
88. RECALL Prediction Positive Negative Reality True True positive False negative False False Positive True negative Recall = True positive True positive + False negative
89. “WITH RECALL, I CAN TELL YOU MOST OF THE POOR WILL DIE”
90. “You can’t be good at precision and recall. Only Chuck Norris can.”
91. How to validate a trained model
92. TRAIN & TEST X y X y y_predict Δ TEST (30%) TRAIN (70%) random
93. CROSS VALIDATION X y X y y_predict Δ TEST (30%) TRAIN (70%) x5 random
94. PREDICT X y_predict NEW DATA
95. RESOURCES • Coursera Machine Learning https://www.coursera.org/learn/machine-learning • Fondamentaux et Etudes de cas, Eric Biernat http://www.amazon.fr/dp/2212142439 • Meetup Machine Learning Paris http://www.meetup.com/fr-FR/Paris-Machine-learning-applications-group/ • Le Meilleur Data Scientist de France http://www.meetup.com/fr-FR/FrenchData/events/228508819/
96. ANY QUESTIONS ? fabien.vauchelles@zelros.com / zelros.com / @fabienv http://bit.ly/scmachinelearning