Exploratory: 決定木の紹介と使い方

機械学習のアルゴリズムのうちの一つで有名な決定木の紹介と、Exploratoryの中での使い方の紹介。

Transcript

1. EXPLORATORY
2. 2 εϐʔΧʔ ੢ా צҰ࿠ CEO EXPLORATORY ུྺ 2016೥ɺσʔλαΠΤϯεͷຽओԽͷͨΊɺExploratory, Inc Λཱͪ ্͛Δɻ Exploratory, Inc.ͰCEOΛ຿ΊΔ͔ͨΘΒɺσʔλαΠΤϯεɾϒʔ τΩϟϯϓɾτϨʔχϯάͳͲΛ௨ͯ͠γϦίϯόϨʔͰߦΘΕ͍ͯ Δ࠷ઌ୺ͷσʔλαΠΤϯεͷීٴͱڭҭʹऔΓ૊Ήɻ ถΦϥΫϧຊࣾͰɺ16೥ʹΘͨΓσʔλαΠΤϯεͷ։ൃνʔϜΛ཰ ͍ɺػցֶशɺϏοάɾσʔλɺϏδωεɾΠϯςϦδΣϯεɺσʔ λϕʔεʹؔ͢Δ਺ଟ͘ͷ੡඼ΛੈʹૹΓग़ͨ͠ɻ @KanAugust
3. Vision ΑΓΑ͍ҙࢥܾఆΛ͢ΔͨΊʹ σʔλΛ࢖͏͜ͱ͕౰ͨΓલʹͳΔ
4. Mission σʔλαΠΤϯεͷຽओԽ
5. 5 ୈ̏ͷ೾ σʔλαΠΤϯεɺAIɺػցֶश͸౷ܭֶऀɺ։ൃऀͷͨΊ͚ͩͷ΋ͷͰ͸͋Γ·ͤΜɻ σʔλʹڵຯͷ͋ΔਓͳΒ୭΋͕ੈքͰ࠷ઌ୺ͷΞϧΰϦζϜΛ࢖ͬͯ ϏδωεσʔλΛ؆୯ʹ෼ੳͰ͖Δ΂͖Ͱ͢ɻ Exploratory͕ͦ͏ͨ͠ੈքΛՄೳʹ͠·͢ɻ
6. ୈ1ͷ೾ ୈ̎ͷ೾ ୈ̏ͷ೾ ϓϥΠϕʔτ(ߴ͍/ݹ͍) Φʔϓϯɾιʔε(ແྉ/࠷ઌ୺) UI & ϓϩάϥϛϯά ϓϩάϥϛϯά 2016 2000 1976 ϚωλΠθʔγϣϯ ίϞσΟςΟԽ ຽओԽ ౷ܭֶऀ σʔλαΠΤϯςΟετ Exploratory ΞϧΰϦζϜ Ϣʔβʔɾ ମݧ πʔϧ Φʔϓϯɾιʔε(ແྉ/࠷ઌ୺) UI & ࣗಈԽ ϏδωεɾϢʔβʔ ςʔϚ σʔλαΠΤϯεͷຽओԽ
7. 質問 ExploratoryͰ؆୯ʹͰ͖ΔλεΫ 伝える データアクセス 加⼯ 可視化 機械学習・AI 統計 UI
8. EXPLORATORY ΦϯϥΠϯɾηϛφʔ
9. Analytics ܾఆ໦
10. 10 σʔλ෼ੳͱ͸ ૬ؔɺύλʔϯΛݟ͚ͭΔ͜ͱ
11. 11 څྉ ೥ྸ ৬छ ۈଓ೥਺ ੑผ 10,000 60 Manager 24 Male 3,000 40 Sales Rep 3 Female 11,000 50 Research Director 35 Female 4,000 20 HR Rep 4 Male 5,000 30 HR Rep 5 Female 10,000 45 Manager 20 Female ஌Γ͍ͨ͜ͱ ଐੑσʔλ
12. 12 ՄࢹԽʂ
13. څྉ vs. ৬छ
14. څྉ vs. ۈଓ೥਺
15. څྉ vs. ֊ڃ
16. 16 σʔλ ૬ؔɾ ύλʔϯ ՄࢹԽͯ͠૬ؔɾύλʔϯΛҰͭҰͭ໨Ͱݟͯݕ஌͢Δ
17. 17 ΊΜͲ͍͘͞ʂ
18. 18 ΞφϦςΟΫεʂ
19. 19 σʔλ ૬ؔɾ ύλʔϯ ػցֶशɾ౷ܭ ΞφϦςΟΫεΛ࢖ͬͯ૬ؔɾύλʔϯΛޮՌతʹݟ͚ͭΔɻ ΞφϦςΟΫε
20. 20 ܾఆ໦ Ϟσϧ ༧ଌϞσϧΛ࡞Δ σʔλ ΞϧΰϦζϜ
21. 21 Monthly Income Age Job Role Department Gender ? 60 Manager Sales Male ? 40 Sales Rep R&D Female ? 30 Research Director HR Female Monthly Income Age Job Role Department Gender 10,000 60 Manager HR Male 11,000 40 Research Director R&D Female 4,000 30 HR Rep HR Female ༧ଌ͢Δ ౴͑ͷͳ͍σʔλ ܾఆ໦ Ϟσϧ
22. 22 Ϟσϧ͸σʔλͷதʹ͋ΔύλʔϯΛ΋ͱʹ࡞ΒΕΔ ܾఆ໦ Ϟσϧ
23. 23 Ͳͷม਺͕ΑΓ૬͕ؔ͋Δͷ͔ɺͲ͏͍͏ؔ܎ੑ Λ͍࣋ͬͯΔͷ͔Λ஌͍ͬͯΔɻ ܾఆ໦ Ϟσϧ
24. 24 σʔλ ΞφϦςΟΫεʹΑͬͯಘΒΕͨΠϯαΠτΛ ՄࢹԽ͢Δ͜ͱͰɺ௚ײతʹཧղ͢Δ ΞφϦςΟΫε ʢػցֶशɺ౷ܭʣ ૬ؔ / ύλʔϯ
25. 25 ܾఆ໦ʢDecision Treeʣ ܾఆ໦͸ɺҰ࿈ͷ࣭໰ͱɺ ͦͷ౴͑ʹΑΔ෼ذͰ݁Ռ Λ༧ଌ͢Δख๏Ͱ͋Δɻ
26. 26 Baby ଻ࣇͷମॏ ଻ࣇͷ਺ ૣ࢈͔Ͳ͏͔ A 5.2 1 TRUE B 4.7 2 TRUE C 6.8 1 FALSE D 7.2 1 FALSE E 5.1 2 TRUE Z 5.8 1 ? ͜ͷ੺ͪΌΜ͸ૣ࢈ʹͳΔͩΖ͏͔ʁ
27. 27 ૣ࢈Λ༧ଌ͢Δܾఆ໦Λ࡞੒ͯ͠ΈΔɻ
28. 28 ଻ࣇͷମॏͱ਺ͷؔ܎ΛՄࢹԽͯ͠ΈΔɻ 28 ଻ࣇͷ਺ ଻ࣇͷମॏ 1 5 2 3 4 5 6 4 7
29. 29 29 ૣ࢈͔Ͳ͏͔ɺͰ৭෼͚Λ͢Δɻ ੺͕ૣ࢈ɺ੨͸ૣ࢈Ͱ͸ͳ͍Λҙຯ͢Δɻ ଻ࣇͷ਺ ଻ࣇͷମॏ 1 5 2 3 4 5 6 4 7
30. 30 30 ௚ઢΛҾ͘͜ͱͰͳΔ΂͘ಉ͡৭ಉ࢜Λάϧʔϓʹ෼͚Δɻ ௚ઢΛҾ͘ճ਺Λ࠷খʹ͢Δ͜ͱΛߟ͑Δɻ ଻ࣇͷ਺ ଻ࣇͷମॏ 1 5 2 3 4 5 6 4 7
31. 31 31 ଻ࣇͷ਺ ଻ࣇͷମॏ 1 5 2 3 4 5 6 4 7 ·ͣɺ଻ࣇͷମॏ͕5.5Ҏ্͔Ͳ͏͔ɺͰάϧʔϓ෼͚Ͱ͖Δɻ ଻ࣇͷମॏ >= 5.5
32. 32 32 ଻ࣇͷ਺ ଻ࣇͷମॏ 1 5 2 3 4 5 6 4 7 ࣍ʹɺ଻ࣇͷ਺͕1.5ΑΓଟ͍͔ɺͰେ͖͘άϧʔϓ෼͚Ͱ͖Δɻ ଻ࣇͷମॏ >= 5.5 ଻ࣇͷ਺ʼ1.5
33. 33 33 ଻ࣇͷ਺ ଻ࣇͷମॏ 1 5 2 3 4 5 6 4 7 ଻ࣇͷମॏ >= 5.5 ଻ࣇͷ਺ʼ1.5 ૣ࢈Ͱ͋Δ: Yes ૣ࢈ͷׂ߹: 100%
 શମͷׂ߹: 40% ૣ࢈Ͱ͋Δ: No ૣ࢈ͷׂ߹: 0% શମͷׂ߹: 40% ૣ࢈Ͱ͋Δ: No ૣ࢈ͷׂ߹: 40% શମͷׂ߹: 20%
34. 34 ଻ࣇͷମॏ >= 5.5 TRUE FALSE ଻ࣇͷ਺ > 1.5 TRUE FALSE 0% 40% 100% ૣ࢈Ͱ͋Δ֬཰
35. 35 Ͳ͏΍ͬͯ໦Λ࡞͍ͬͯΔͷ͔
36. 36 Ͳͷ࣭໰ʢ৚݅ʣΛઌʹ࣋ͬͯ͘Δ͔
37. 37 ෆ७౓ʢGini Impurityʣ • 0͔Β1ͷؒͷ஋ΛऔΔɻ • ͦΕͧΕͷϊʔυͷσʔλʹͲΕ͚ͩ஋͕ࠞ ͍ͬͯ͟Δ͔Λද͢ࢦඪ
38. pi 38 ෆ७౓ʢGini Impurityʣ ෆ७౓ (Gini Impurity) ͸ͦͷϊʔυʹ͋ΔҰҙͷ஋ͷ਺Λnͱ͢Δͱɺ ҎԼͷΑ͏ʹܭࢉͰ͖Δɻ( ͸i൪໨ͷ஋Λ࣋ͭαϯϓϧ਺ͷׂ߹) 1 − p2 1 − p2 2 − p2 3 − . . . . p2 n
39. ෆ७౓ = 0 39 Not ૣ࢈ Not ૣ࢈ Not ૣ࢈ 1 - (0/6)2 - (6/6)2 = 0 Not ૣ࢈ Not ૣ࢈ Not ૣ࢈
40. ෆ७౓ = 0 40 ૣ࢈ ૣ࢈ ૣ࢈ 1 - (6/6)2 - (0/6)2 = 0 ૣ࢈ ૣ࢈ ૣ࢈
41. ෆ७౓ = 0.44 41 Not ૣ࢈ Not ૣ࢈ Not ૣ࢈ ૣ࢈ ૣ࢈ 1 - (2/6)2 - (4/6)2 = 0.44 Not ૣ࢈
42. ෆ७౓ = 0.44 42 Not ૣ࢈ ૣ࢈ ૣ࢈ 1 - (4/6)2 - (2/6)2 = 0.44 Not ૣ࢈ ૣ࢈ ૣ࢈
43. ෆ७౓ = 0.5 43 Not ૣ࢈ ૣ࢈ Not ૣ࢈ Not ૣ࢈ ૣ࢈ ૣ࢈ 1 - (3/6)2 - (3/6)2 = 0.5
44. 44 ૣ࢈ Not ૣ࢈ ૣ࢈ ૣ࢈ Impurity: 0.5 ૣ࢈ ૣ࢈ Not ૣ࢈ Not ૣ࢈ Not ૣ࢈ Not ૣ࢈ ελʔτ஍఺
45. 45 ଻ࣇͷମॏͰ࠷ॳʹάϧʔϓ෼͚͢Δ৔߹
46. 46 ଻ࣇͷମॏ >= 5.5 TRUE FALSE
47. 47 ଻ࣇͷମॏ >= 5.5 TRUE FALSE ෆ७౓: 0 ෆ७౓: 1- (2/7)2 - (5/7)2 = 0.41
48. 48 ଻ࣇͷମॏ >= 5.5 TRUE FALSE ෆ७౓: 0 ෆ७౓: 1- (2/7)2 - (5/7)2 = 0.41 ෆ७౓: 3/10*0 + 7/10*0.41 = 0.29
49. 49 ଻ࣇͷମॏ >= 5.5 TRUE FALSE ෆ७౓: 3/10*0 + 7/10*0.41 = 0.29 ෆ७౓: 0.5
50. 50 ଻ࣇͷମॏ >= 5.5 TRUE FALSE ෆ७౓: 3/10*0 + 7/10*0.41 = 0.29 ෆ७౓: 0.5 ෆ७౓ͷݮগ: 0.21
51. 51 ଻ࣇͷ਺Ͱ࠷ॳʹάϧʔϓ෼͚͢Δ৔߹
52. 52 ଻ࣇͷ਺ > 1.5 TRUE FALSE
53. 53 ଻ࣇͷମॏ >= 5.5 TRUE FALSE ଻ࣇͷ਺ > 1.5 ෆ७౓: 1- (2/5)2 - (3/5)2 = 0.48 ෆ७౓: 1- (3/5)2 - (2/5)2 = 0.48
54. 54 ଻ࣇͷମॏ >= 5.5 TRUE FALSE ෆ७౓: 5/10*0.48 + 5/10*0.48 = 0.48 ଻ࣇͷ਺ > 1.5 ෆ७౓: 1- (2/5)2 - (3/5)2 = 0.48 ෆ७౓: 1- (3/5)2 - (2/5)2 = 0.48
55. 55 ଻ࣇͷମॏ >= 5.5 TRUE FALSE ෆ७౓: 5/10*0.48 + 5/10*0.48 = 0.48 ଻ࣇͷ਺ > 1.5 ෆ७౓: 0.5
56. 56 ଻ࣇͷମॏ >= 5.5 TRUE FALSE ෆ७౓: 5/10*0.48 + 5/10*0.48 = 0.48 ଻ࣇͷ਺ > 1.5 ෆ७౓: 0.5 ෆ७౓ͷݮগ: 0.02
57. 57 ଻ࣇͷମॏͷํ͕ɺ ଻ࣇͷ਺ΑΓෆ७౓ΛݮΒͤΔ ଻ࣇͷ਺ 0.02 ʻ 0.48 ଻ࣇͷମॏ ෆ७౓ͷݮগΛൺֱ͢Δͱŋŋŋ
58. 58 ઌʹ଻ࣇͷମॏͰάϧʔϓ෼͚ͯ͠ ࣍ʹ଻ࣇͷ਺Ͱάϧʔϓ෼͚͢Δ ໦Λ࡞Δ
59. 59 TRUE FALSE TRUE FALSE 100% 50% 0% ଻ࣇͷମॏ >= 5.5 ଻ࣇͷ਺ > 1.5
60. 60 ͢Δͱɺͦ͏Ͱͳ͍৔߹ʹൺ΂ͯ গͳ͍࣭໰ʢ෼ذʣͰ͢Ή
61. 61 ઌʹ଻ࣇͷ਺Ͱάϧʔϓ෼͚ͯ͠ ࣍ʹ଻ࣇͷମॏͰάϧʔϓ෼͚͢Δ ໦Λ࡞Δ
62. 62 Over_35 TRUE FALSE Is_Plural TRUE FALSE 50% 100% Is_Plural TRUE FALSE 100% 100% ଻ࣇͷମॏ >= 5.5 ଻ࣇͷ਺ > 1.5 ଻ࣇͷମॏ >= 5.5
63. 63 ෼ྨ vs. ճؼ
64. 64 ੜ·Εͯ͘Δ੺ͪΌΜ͸ະख़ࣇ͔ʁ ෼ྨ ෼ྨ vs. ճؼ ճؼ ͍ͭ੺ͪΌΜ͕ੜ·Εͯ͘Δ͔ʁ
65. 65 65 ෼ྨ ଻ࣇͷ਺ ଻ࣇͷମॏ 1 5 2 3 4 5 6 4 7
66. 66 Mother Age Father Age ճؼ
67. 67 Is_Prural TRUE FALSE Over_35 TRUE FALSE ೛৷ظؒΛ༧ଌ 20 40
68. 68 Is_Prural TRUE FALSE Over_35 TRUE FALSE 20 weeks 30 weeks 37 weeks ೛৷ظؒΛ༧ଌ 20 ฏۉ 40 ฏۉ͔Βͷ͹Β͖͕ͭ ࠷খʹͳΔΑ͏ʹ ࢬΛ෼͚͍ͯ͘
69. 69 ΞφϦςΟΫε ܾఆ໦Λ࢖ͬͯݟΔʂ
70. is_premature(ૣ࢈)ͷྻΛ࡞Δ ΋͠ɺis_prematureྻ͕ͳ͍৔߹ɺૣ࢈͔Ͳ͏͔ (37िະ ຬ͔Ͳ͏͔)ͷ࿦ཧ஋ΛͱΔྻΛgestation_weeks(೛৷ि) ͷྻ͔Β৽ͨʹ࡞Δɻ 70 gestation_weeks < 37
71. is_prematureྻΛ࡞Δ gestation_weeks(೛৷ि)ͷྻϔομϝχϡʔ͔Βܭࢉͷ࡞੒(Mutate)ΛબͿɻ 71
72. 72 • ྻ໊ʹ is_prematureͱೖྗɻ • ܭࢉʹ gestation_weeks<37 Λೖྗɻ ࡞੒͞ΕΔྻʹ͸ɺ37िະຬͳΒTRUEɺ 37िҎ্ͳΒFALSEͷ஋͕ೖΔɻ is_prematureྻΛ࡞Δ
73. 73 ܾఆ໦ΞφϦςΟΫε
74. 74 ༧ଌର৅ྻͷબ୒
75. 75 ม਺ͷྻͷબ୒
76. 76 gestation_weeks(೛৷ि)Ҏ֎ͷશͯͷྻΛબ୒
77. 77 ܾఆ໦͕࡞੒͞Εͨɻ
78. ։࢝஍఺
79. ଟ਺೿ͷσʔλ : FALSE (Not ૣ࢈). TRUE (Premature) ͷׂ߹ : 12%. ͜ͷ஍఺Ͱͷσʔλͷׂ߹ : 100% ։࢝஍఺
80. ৚݅: ମॏʢweight_pounds greaterʣ͕ 5.3 ύ΢ϯυҎ্͔?
81. ଟ਺೿ͷσʔλɿFALSE TRUE (Premature) ͷׂ߹ : 8%. ͜ͷ஍఺Ͱͷσʔλͷׂ߹ : 94%
82. ଟ਺೿ͷσʔλɿTRUE TRUE (Premature) ͷׂ߹ : 72%. ͜ͷ஍఺Ͱͷσʔλͷׂ߹ : 6%
83. σʔλΛՄࢹԽ͔ͯ֬͠ΊΔ
84. Is_Premature vs. Weight
85. Is_Premature vs. Weight
86. ैۀһσʔλΛ࢖ͬͨྫ
88. ܾఆ໦ͷϞσϧΛ࡞Δ
91. σʔλΛՄࢹԽ͔ͯ֬͠ΊΔ
92. Attrition vs. Overtime
93. Attrition vs. Monthly Income
95. • ϓϩάϥϛϯάͳ͠ RݴޠͷUIͰ͋ΔExploratoryΛ෼ੳπʔϧͱͯ͠࢖༻͢ΔͨΊडߨத͸ɺϏδωεͷ ໰୊Λղܾ͢ΔͨΊʹඞཁͳσʔλαΠΤϯεͷख๏ͷशಘʹ100ˋूதͰ͖Δ • ෼ੳπʔϧͷϕϯμʔϩοΫΠϯͳ͠ ExploratoryͰͷ࡞ۀ͸શͯಠཱͨ͠ΦʔϓϯιʔεͷR؀ڥͰ࠶ݱ͕Մೳ • ࢥߟྗͱεΩϧͷशಘ σʔλαΠΤϯεͷεΩϧशಘ͚ͩͰͳ͘ɺσʔλ෼ੳʹඞཁͳࢥߟྗ΋शಘͰ͖Δ ಛ௃
96. Q & A
97. ࿈བྷઌ ϝʔϧ kan@exploratory.io ΢ΣϒαΠτ https://ja.exploratory.io ϒʔτΩϟϯϓɾτϨʔχϯά https://ja.exploratory.io/training-jp Twitter @KanAugust