For many tasks, it makes little difference if these programs are opaque to human introspection. Here, high capacity models, like deep learning, suffer little penalty for representational complexity.
However, for several reasons, marketers tend to be wary about ceding control of their customers’ experiences to black box methods.
This presentation covers Conductrics approach to generating machine learning for marketing optimization that is both machine and human readable.
NO1 Certified Black Magic Specialist Expert Amil baba in Lahore Islamabad Raw...
Conductrics interpretable Machine Learning
1.
2. Agenda
1. Quick overview of Machine learning
2. Complexity is costly
3. Interpretability vs Accuracy
4. AB Testing is NOT Optimization
5. This is hard – which is Good!
6. Get drinks
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
4. Machine Learning Definition
Statistics Computer
Science
see: Mitchel http://www.cs.cmu.edu/~tom/pubs/MachineLearning.pdfConductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
5. Machine Learning Definition
Statistics Computer
Science
Machine
Learning
see: Mitchel http://www.cs.cmu.edu/~tom/pubs/MachineLearning.pdfConductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
6. Machine Learning Definition
Methods that generate useful
computer programs via
interaction with Data or the
Environment
see: Mitchel http://www.cs.cmu.edu/~tom/pubs/MachineLearning.pdf
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
7. Example of Machine Learning
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
24. Online Optimization
Pixel Data Customer Data
• Weekend/Weekday
• Mobile/Desktop
• Browser Type
• User Age
• Geo/Census
• Weather
• Tenure/RFM Score
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
29. Arguments for Interpretability
1. Trust
2. Insights
3. Review/Accountability
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
30. Arguments for Interpretability
1. Trust
2. Insights
3. Review/Accountability
4. Communication
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
31. General Data Protection Regulation
1. EU Data Protection Laws
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
32. General Data Protection Regulation
1. EU Data Protection Laws
2. Goes in affect May 2018
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
33. General Data Protection Regulation
1. EU Data Protection Laws
2. Goes in affect May 2018
3. Article 22
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
35. GDPR: Article 22
Automated Decisions are Contestable
1. What data was used?
2. Explain why a decision was made
to EU citizens.
3. Non discrimination
4. Up to 4% world gross revenues
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
36. What are some Options?
1. Use Complex Model and
then try to interpret/explain
2. Use Simple interpretable
model
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
38. L Local
I Interpretable
M Model-Agnostic
E Explanations
Lime Interpreting Complex Models
ACM SIGKDD International Conference on Knowledge Discovery and Data Mining
(KDD), 2016 Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
42. 1.Works for Classification problems (not
regression yet)
2.Approximate Explanations (Lossy)
3.Explanation is not the model that is
actually used – maybe an issue for GDPR?
Lime: Interpreting Complex Models
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
43. Don’t Build Complex Model
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
44. Atari – Lots of Structure
By Jorge Stolfi (Own work) [CC BY-SA 3.0 (http://creativecommons.org/licenses/by-sa/3.0)], via Wikimedia Commons
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
45. Marketing Problems
By Jorge Stolfi (Own work) [CC BY-SA 3.0 (http://creativecommons.org/licenses/by-sa/3.0)], via Wikimedia Commons
• High Noise
• Often Low(ish) N
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
46. Simple Models
Simple Models have Lower Opportunity Costs
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
47. Simple Models
Simple Models have Lower Opportunity Costs
IF World is
By Jorge Stolfi (Own work) [CC BY-SA 3.0 (http://creativecommons.org/licenses/by-sa/3.0)], via Wikimedia Commons
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
48. Simple Models
Simple Models have Lower Opportunity Costs
IF World is
Rather Than
By Jorge Stolfi (Own work) [CC BY-SA 3.0 (http://creativecommons.org/licenses/by-sa/3.0)], via Wikimedia Commons
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
49. What Simple Model?
Linear Model
𝑌 = 𝑋𝛽 + 𝜖
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
50. Linear Model – the good
𝑌 = 𝑋𝛽 + 𝜖
1. Online / SGD
What Simple Model?
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
51. 𝑌 = 𝑋𝛽 + 𝜖
1. Online / SGD
2. ’s interpretable
Linear Model – the good
What Simple Model?
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
52. 𝑌 = 𝑋𝛽 + 𝜖
1. Online / SGD
2. ’s interpretable
3. BLUE (sort of)
Linear Model – the good
What Simple Model?
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
53. Linear Model – the Bad
𝑌 = 𝑋𝛽 + 𝜖
What Simple Model?
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
54. Linear Model – the Bad
𝜷 𝐚𝟎 ⋯ 𝜷 𝐚𝒎
⋮ ⋱ ⋮
𝜷 𝒌𝟎 ⋯ 𝜷 𝒌𝒎
𝑌 = 𝑋𝛽 + 𝜖
What Simple Model?
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
55. Linear Model – the Bad
𝜷 𝐚𝟎 ⋯ 𝜷 𝐚𝒎
⋮ ⋱ ⋮
𝜷 𝒌𝟎 ⋯ 𝜷 𝒌𝒎
What Simple Model?
Still Not really that interpretable
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
56. Decision Tree – Better!
What Simple Model?
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
57. Decision Tree – Has nice properties
Use Simple Model?
1)Human Readable for Trust and Insights
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
58. Decision Tree – Has nice properties
Use Simple Model?
1)Human Readable for Trust and Insights
2)Easily to Audit for review
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
59. Decision Tree – Has nice properties
Use Simple Model?
1)Human Readable for Trust and Insights
2)Easily to Audit for review
3)Loggable
1)Decision policy is represented as rules
2)Compact
3)Each decision can be logged with the exact
policy/rule
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
62. Decision Tree from Model
Use a fact of Regression +
a little algebra
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
63. Decision Tree from Model
𝑌 = ത𝑌 +
𝑗
𝑗 ∗ 𝑥𝑗 − ഥ𝑥𝑗
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
64. Decision Tree from Model
• ‘Set’ all targeting to ‘0’
𝑌 = ത𝑌 + 0
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
65. Decision Tree from Model
𝑌 = ത𝑌
• ‘Set’ all targeting to ‘0’
• Prediction reduces to Mean
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
66. 1. Start with Simple Means (A,B,C, etc.)
Decision Tree from Model
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
67. Decision Tree from Model
Start Root Node Using unconditional
means
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
68. 1. Start with Simple Means (A,B, etc.)
2. Collect Data ->Learn Model
Decision Tree from Model
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
69. 1. Start with Simple Means (A,B,C, etc.)
2. Collect Data ->Learn Model
3. Play ‘20’ Questions w/ model
Decision Tree from Model
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
70. 1. Start with Simple Means (A,B,C, etc.)
2. Collect Data ->Learn Model
3. Play ‘20’ Questions w/ model
4. Add Features to tree
Decision Tree from Model
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
71. 1. Start with Simple Means (A,B,C, etc.)
2. Collect Data ->Learn Model
3. Play ‘20’ Questions w/ model
4. Add Features to tree
5. Update Simple Means using this equation
𝑌 = ത𝑌 +
𝑗
𝑗 ∗ 𝑥𝑗 − ഥ𝑥𝑗
Decision Tree from Model
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
75. 1.Online Learning / Batch Controller
Benefits for Conductrics
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
76. 1.Online Learning / Batch Controller
2.Degrade gracefully – so shrink back to means
Benefits for Conductrics
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
77. 1.Online Learning / Batch Controller
2.Degrade gracefully – so shrink back to means
3.Shrinkage/partial pooling on means
Benefits for Conductrics
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
78. 1.Online Learning / Batch Controller
2.Degrade gracefully – so shrink back to means
3.Shrinkage/partial pooling on means
4.Add arbitrary constraints:
1.Size of leaf nodes
2.Depth of tree
3.Max Leaves
Benefits for Conductrics
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
80. 1) Importance of Interpretabilty
2) Marketing Problems often Noisy
3) Decision Trees good representation
What did we Do/Learn?
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
81. 1) Importance of Interpretabilty
2) Marketing Problems often Noisy
3) Decision Trees good representation
What did we Do/Learn?
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics
82. Learn more!
1. www.conductrics.com
2. https://conductrics.com/machine-learning-and-human-interpretability
3. info@conductrics.com
Conductrics Inc. | Matt Gershoff |
www.conductrics.com | @conductrics