The document discusses several lessons learned about machine learning including: 1) Sometimes more data is not needed and better models can be achieved through feature engineering and appropriate hyperparameters. 2) Models will learn what they are taught through the training data, so training data needs to be thoughtful and free of biases. 3) A combination of supervised and unsupervised learning along with ensemble methods often provides the best results rather than relying on any single approach.

1. LessonsLearned

9. LessonsLearned

10. MoreDatavs.BetterModels

11. Really?
Anand Rajaraman: Former Stanford Prof. & Senior VP at Walmart

12. Sometimes, it’s
not about more
data

13. Norvig: “Google
does not have
better Algorithms
only more Data”
Many features/
low-bias models

14. Sometimes, it’s
not about more
data

15. YouMightnotneed
allyour“bigData”

16. ○
○

17. Sometimesyoudoneed
aComplexModel

19. Itpaysofftobesmartabout
Hyperparameters

20. ○
○

21. Supervisedvs.plus
UnsupervisedLearning

22. ○
○
○
○
○

23. ○
○

24. Everythingisanensemble

25. ○
○
○
○
○
○

26. ○
○

28. Theoutputofyourmodel
willbetheinputofanotherone
(andothersystemdesignproblems)

29. ○
○
○

30. ○

31. Thepains&gains
ofFeatureEngineering

32. ○
○
○
○

33. ○
○
○
○

36. Implicitsignalsbeat
explicitones
(almostalways)

38. ○
○
○
○

39. ○
○
○

40. bethoughtfulaboutyour
TrainingData

41. ○
○

42. ○
○
○

43. YourModelwilllearn
whatyouteachittolearn

44. ○
○
○
○

45. ○

46. Learntodealwith
PresentationBias

47. More likely
to see
Less likely

49. DataandModelsaregreat.Youknowwhat’
sevenbetter?
Therightevaluationapproach!

51. ○
○

53. Youdon’tneedtodistribute
yourMLalgorithm

54. ○
○
○

55. ○
○
○

56. but,Ifyoudo,youshouldunderstandat
whatleveltodoit

57. The three levels of Distribution/Parallelization
● For each subset of the population (e.g. region)
● For each combination of the hyperparameters
● For each subset of the training data
Each level has different requirements
ANN Training over distributed GPU’s

58. somethingsarebetterdone Online and
othersoffline…and,thereis Nearlinefor
everythinginbetween

59. System Overview
● Blueprint for multiple
personalization algorithm
services
● Ranking
● Row selection
● Ratings
● …
● Recommendation involving
multi-layered Machine
Learning

60. Matrix Factorization Example

61. Thetwofacesofyour
MLinfrastructure

62. ○
○

63. ○
○
○
○

64. ○
○
○
○

65. ○
○

66. Whyyoushouldcareabout
answeringquestions(aboutyourmodel)

67. ○
○
○

69. Theuntoldstoryof
DataScienceandvs.MLengineering

70. ○
○
○

72. ○
○

74. ○
○