Re:Predicting the Oscars with Machine Learning
by aficionado on Tuesday January 15, 2013 @12:21AM
(#5058803)
I used Megafauna Software parsed data to quickly create a model using BigML: http://bml.io/W6lKqe It uses information gain mix to create a decision tree. It also concluded that winning the 'best director' is the most important field. However, 'make up' didn't top the most important fields. See the results below.
Data distribution:
n: 16.94% (82 instances)
y: 83.06% (402 instances)
Predicted distribution:
n: 16.74% (81 instances)
y: 83.26% (403 instances)
Field importance:
1. Directing_1: 84.47%
2. Music (Scoring)_0: 3.59%
3. Music (Scoring)_1: 2.68%
4. Art Direction_0: 2.39%
5. Sound_0: 1.75%
6. Film Editing_0: 1.55%
7. Costume Design_1: 1.40%
8. Film Editing_1: 0.82%
9. Costume Design_0: 0.54%
10. Cinematography_1: 0.52%
11. Art Direction_1: 0.29%
12. Special Achievement Award_0: 0.00%
13. Scientific and Technical (Bonner Medal)_1: 0.00%
14. Foreign Language Film_0: 0.00%
15. Sound Editing_1: 0.00%
16. Scientific and Technical (Bonner Medal)_0: 0.00%
17. Sound Editing_0: 0.00%
18. Scientific and Technical (Academy Award of Merit)_1: 0.00%
19. Sound_1: 0.00%
20. Animated Feature Film_1: 0.00%
21. Scientific and Technical (Academy Award of Merit)_0: 0.00%
22. Documentary (other)_1: 0.00%
23. Animated Feature Film_0: 0.00%
24. Music (Song)_1: 0.00%
25. Documentary (other)_0: 0.00%
26. Actress -- Supporting Role_1: 0.00%
27. Short Film (Live Action)_1: 0.00%
28. Music (Song)_0: 0.00%
29. Documentary (Short Subject)_1: 0.00%
30. Actress -- Supporting Role_0: 0.00%
31. Short Film (Live Action)_0: 0.00%
32. Documentary (Short Subject)_0: 0.00%
33. Actress -- Leading Role_1: 0.00%
34. Short Film (Animated)_1: 0.00%
35. Documentary (Feature)_1: 0.00%
36. Short Film (Animated)_0: 0.00%
37. Makeup_1: 0.00%
38. Scientific and Technical (Technical Achievement Award)_1: 0.00%
39. Actress -- Leading Role_0: 0.00%
40. Documentary (Feature)_0: 0.00%
41. Actor -- Supporting Role_1: 0.00%
42. Makeup_0: 0.00%
43. Actor -- Supporting Role_0: 0.00%
44. Scientific and Technical (Technical Achievement Award)_0: 0.00%
45. Jean Hersholt Humanitarian Award_1: 0.00%
46. Directing_0: 0.00%
47. Actor -- Leading Role_1: 0.00%
48. Scientific and Technical (Special Awards)_1: 0.00%
49. Jean Hersholt Humanitarian Award_0: 0.00%
50. Actor -- Leading Role_0: 0.00%
51. Writing_1: 0.00%
52. Scientific and Technical (Special Awards)_0: 0.00%
53. Irving G. Thalberg Memorial Award_1: 0.00%
54. Acting (other)_1: 0.00%
55. Writing_0: 0.00%
56. Scientific and Technical (Scientific and Engineering Award)_1: 0.00%
57. Irving G. Thalberg Memorial Award_0: 0.00%
58. Acting (other)_0: 0.00%
59. Visual Effects_1: 0.00%
60. Scientific and Technical (Scientific and Engineering Award)_0: 0.00%
61. Honorary Award_1: 0.00%
62. Cinematography_0: 0.00%
63. Visual Effects_0: 0.00%
64. Scientific and Technical (Gordon E. Sawyer Award)_1: 0.00%
65. Honorary Award_0: 0.00%
66. Category_1: 0.00%
67. Special Achievement Award_1: 0.00%
68. Scientific and Technical (Gordon E. Sawyer Award)_0: 0.00%
69. Foreign Language Film_1: 0.00%
70. Category_0: 0.00%
Re:Predicting the Oscars with Machine Learning
by megafaunasoft on Saturday January 19, 2013 @02:01PM
(#5084337)
Very cool -- I'm glad someone else could use the data!
BigML seems useful. The BigML decision tree looks quite a lot like the J48 tree result. I wonder if the performance (ROC) is closer to J48 than logistic regression. I couldn't see an option to cross-validate on BigML.
Re:Predicting the Oscars with Machine Learning
by kenjackson on Monday January 14, 2013 @11:36AM
(#5055294)
It would be interesting to use some other data to predict the Oscars -- data that comes out before the Oscars begins, e.g., Golden Globe and other film festival awards, date the movie was released, box office revenue, availability on DVD, history of actors/directors/writers (e.g., have they won or been nominated for Oscars in the past).
Re:Predicting the Oscars with Machine Learning
by Breakthrough on Monday January 14, 2013 @12:37PM
(#5055592)
Perhaps this illustrates how much of an effect a film's director can have on it's perception.
As Stanley Kubrick once said, "The screen is a magic medium. It has such power that it can retain interest as it conveys emotions and moods that no other art form can hope to tackle."
BigML seems useful. The BigML decision tree looks quite a lot like the J48 tree result. I wonder if the performance (ROC) is closer to J48 than logistic regression. I couldn't see an option to cross-validate on BigML.
As Stanley Kubrick once said, "The screen is a magic medium. It has such power that it can retain interest as it conveys emotions and moods that no other art form can hope to tackle."