Augmenting Human Decision Making with Data Science - PyCon

Transcript

1. HUMAN DECISION MAKING Kelsey Pedersen Software Engineer, Stitch Fix with DATA SCIENCE AUGMENTING @kelsey_pedersen
2. 72% of Americans are scared of computers taking over our jobs @kelsey_pedersen
3. 51% worry about gun controls
4. 55% worry about affordable health care
5. Gun Controls Health Care Algorithms & Robots 51% 55% 72% @kelsey_pedersen
7. @kelsey_pedersen
8. How to balance human decisions with algorithmic decisions in our software GOAL
9. @kelsey_pedersen KELSEY PEDERSEN
10. @kelsey_pedersen
12. What are the limitations of human decision making? What are the limitations of algorithmic decision making? How can human decisions be augmented by data science? @kelsey_pedersen
13. What are the limitations of human decision making?
14. System 1 System 2
15. Intuition & feelings System 1
16. FAST AUTOMATIC UNCONSCIOUS System 1 @kelsey_pedersen
17. Interpretation of our surroundings System 1
22. 95% of human decisions are made in System 1
23. Analytical & effortful System 2
25. 1737 x 1990 Answer: 3,456,630 @kelsey_pedersen
27. System 1 gut feelings System 2 computations
28. @kelsey_pedersen
29. Gut feelings are unpredictable @kelsey_pedersen
30. Environment & mood Influence thoughts and feelings
31. INCONSISTENT INFO OVERLOAD BIASED
32. Gut feelings are driven by your own views and preferences @kelsey_pedersen
33. Biases occur outside of our own awareness Cause us to think and act irrationally
34. Computation of lots of info takes time and energy @kelsey_pedersen
35. Lots of information Causes physical response
36. INCONSISTENT || BIASED || INFO OVERLOAD
37. Our stylists are human, so these inconsistencies apply to them too. @kelsey_pedersen
38. INCONSISTENT INFO OVERLOAD BIASED
39. Inconsistent Judgments by Stylists
40. Biased Decisions by Stylists
41. Information Overload by Stylists
42. @kelsey_pedersen
43. Case Study @kelsey_pedersen
44. Our data science team uses multiple sources of internal data @kelsey_pedersen Direct from our customers
45. Did a customer keep or return an item? Customer survey * What size top are you? S, M, L, XL Buying Patterns Size of this item? Too small, just right, too big Checkout feedback * Helps the cold start problem Direct Feedback Indirect Feedback
46. FIT @kelsey_pedersen
47. FIT STYLE @kelsey_pedersen
48. FIT STYLE PRICE @kelsey_pedersen
49. FIT STYLE PRICE SIZE @kelsey_pedersen
50. @kelsey_pedersen
51. How can human decisions be augmented by data science?
52. @kelsey_pedersen
53. (1) Stylist deciding on one item
55. 72% 60% 48% 44% @kelsey_pedersen
56. @kelsey_pedersen
57. @kelsey_pedersen
58. @kelsey_pedersen
59. (2) Stylist deciding on all 5 items
60. 72% 60% 48% 44% 56% @kelsey_pedersen
61. 50% 26% 18% 14% 27% @kelsey_pedersen
62. @kelsey_pedersen
63. @kelsey_pedersen
64. (3) One Client’s Feedback on all 5 items
65. @kelsey_pedersen
66. @kelsey_pedersen
68. Stylists use feedback to train to make more accurate decisions in the future @kelsey_pedersen
69. (4) All Clients’ Feedback for one Stylist
70. STATS
72. Stylist Stylist Manager
73. (5) All Clients’ Feedback overtime for all Stylists
76. Deciding on one item Client Feedback on all 5 items Deciding on all 5 items @kelsey_pedersen Client Feedback overtime for one Stylist All Feedback overtime for all Stylists GUIDE TRAIN
77. GUIDE DECISIONS WITH COMPUTATIONS TRAIN DECISIONS WITH FEEDBACK
78. What are the limitations of algorithmic decision making?
82. Stylists are able to override the algorithms. @kelsey_pedersen
84. 72% 60% 48% 10% @kelsey_pedersen
85. 72% 60% 48% 44% 24% @kelsey_pedersen
86. When intuition doesn’t match the algorithm, we can learn from that. @kelsey_pedersen
89. stylists data scientists @kelsey_pedersen
90. What is the future of humans and data science?
91. @kelsey_pedersen
93. System 1 gut feelings System 2 computations &
94. @kelsey_pedersen
95. Data Science Humans System 1 System 2 Data Science Humans System 3 @kelsey_pedersen
96. Predictive & intuitive System 3
98. System 1 gut feelings System 2 computations System 3 feedback
99. What is the business impact of system 3?
100. lower labor cost fewer mistakes from humans greater client satisfaction increased keep rate @kelsey_pedersen
101. In conclusion…
102. Conclusion 95%
103. INCONSISTENT INFO OVERLOAD BIASED
105. Conclusion @kelsey_pedersen
106. Humans lack the ability to process large volumes of information. @kelsey_pedersen
107. Machines lack intuition, empathy, nuance and ethics. @kelsey_pedersen
109. @kelsey_pedersen
110. Conclusion @kelsey_pedersen
111. humans data science @kelsey_pedersen
112. Thank you!
113. @kelsey_pedersen Thanks! KELSEY PEDERSEN Stitch Fix Promo bit.ly/pycon-stitchfix We’re hiring!