Presented at #H2OWorld 2017 in Mountain View, CA. Enjoy the video: https://youtu.be/axIqeaUhow0. Learn more about H2O.ai: https://www.h2o.ai/. Follow @h2oai: https://twitter.com/h2oai. - - - Abstract: Usage of AI and machine learning models is likely to become more commonplace as larger swaths of the economy embrace automation and data-driven decision-making. While these predictive systems can be quite accurate, they have been treated as inscrutable black boxes in the past, that produce only numeric predictions with no accompanying explanations. Unfortunately, recent studies and recent events have drawn attention to mathematical and sociological flaws in prominent weak AI and ML systems, but practitioners usually don’t have the right tools to pry open machine learning black-boxes and debug them. This presentation introduces several new approaches to that increase transparency, accountability, and trustworthiness in machine learning models. If you are a data scientist or analyst and you want to explain a machine learning model to your customers or managers (or if you have concerns about documentation, validation, or regulatory requirements), then this presentation is for you!

2. Patrick Hall, Navdeep Gill, Mark Chan

3. Machine Learning Interpretability (MLI)

4. MEET
THE MAKERS
PATRICK HALL MARK CHANNAVDEEP GILL
• Patrick Hall is a senior director for data science products at
H2O.ai and adjunct faculty in the Department of Decision
Sciences at George Washington University. He is the lead author
of a popular white paper on techniques for interpreting machine
learning models and a frequent speaker on the topics of FAT/ML
and explainable artificial intelligence (XAI) at conferences and on
webinars.
• Navdeep Gill is a software engineer and data scientist at H2O.ai.
He has made important contributions to the popular open source
h2o machine learning library and the newer open source h2o4gpu
library. Navdeep also led a recent Silicon Valley Big Data Science
Meetup about interpretable machine learning.
• Mark Chan is a software engineer and customer data scientist at
H2O.ai. He has contributed to the open source h2o library and to
critical financial services customer products.

5. First-time Qwiklab Account Setup
• Go to http://h2oai.qwiklab.com
• Click on “JOIN” (top right)
• Create a new account with a valid email address
• You will receive a confirmation email
• Click on the link in the confirmation email
• Go back to http://h2oai.qwiklab.com and log in
• Go to the Catalog on the left bar
• Choose “Introduction to Driverless AI”
• Wait for instructions

6. MLI in Academia/Press

7. Big Ideas

8. Learning from data …
Adapted from:
Learning from Data. https://work.caltech.edu/textbook.html

9. EXPLAIN
HYPOTHESIS
h ≈ g, βj g(x(i)
j), g(x(i)
(-j))
(explain predictions with reason codes)
Learning from data …
transparently.
Adapted from:
Learning from Data. https://work.caltech.edu/textbook.html

10. Increasing fairness, accountability, and trust by
decreasing unwanted sociological biases
Source: http://money.cnn.com/, Apple Computers

11. A framework for interpretability
Complexity of learned functions:
• Linear, monotonic
• Nonlinear, monotonic
• Nonlinear, non-monotonic
(~ Number of parameters/VC dimension)
Enhancing trust and understanding:
the mechanisms and results of an
interpretable model should be both
transparent AND dependable.
Understanding ~ transparency
Trust ~ fairness and accountability
Scope of interpretability:
Global vs. local
Application domain:
Model-agnostic vs. model-specific

12. Big Challenges

13. Linear Models
Strong model locality
Usually stable models and
explanations
Machine Learning
Weak model locality
Sometimes unstable models and
explanations
(a.k.a. The Multiplicity of Good Models )

14. Age
Number of Purchases
Lost profits.
Wasted marketing.
“For a one unit increase in age, the number
of purchases increases by 0.8 on average.”
Linear Models
Machine Learning
Exact explanations
for approximate
models.
Approximate
explanations for exact
models.
Age
“Slope begins to
decrease here. Act to
optimize savings.”
“Slope begins to
increase here sharply.
Act to optimize profits.”
Number of Purchase

15. A Few of Our Favorite Things

16. Partial dependence plots
Source: http://statweb.stanford.edu/~tibs/ElemStatLearn/printings/ESLII_print10.pdf
HomeValue ~ MedInc + AveOccup + HouseAge + AveRooms

17. Surrogate models

18. Local interpretable model-agnostic
explanations (LIME)
Source: https://github.com/marcotcr/lime
Weighted
explanatory
samples.
Linear model used to explain
nonlinear decision boundary in
local region.

19. Variable importance measures
Global variable
importance indicates
the impact of a variable
on the model for the
entire training data set.
Local variable
importance can
indicate the impact of a
variable for each
decision a model
makes – similar to
reason codes.

20. Current product roadmap
Time Frame Features
Near-Term Reason Codes in MOJO (i.e. Prod), Sensitivity
Analysis, Multinomial Explanations
Medium-Term Table Plots, Residual Analysis, Python API,
Performance Refactor (GPU), Report Export
Long-Term R API, AutoMLI
(Roadmap subject to change without notice.)

21. Resources

22. Machine Learning Interpretability with
H2O Driverless AI
http://docs.h2o.ai/driverless-ai/latest-stable/docs/booklets/MLIBooklet.pdf
Ideas on Interpreting Machine Learning
https://www.oreilly.com/ideas/ideas-on-interpreting-machine-learning
FAT/ML
http://www.fatml.org/
MLI Resources
https://github.com/h2oai/mli-resources

23. MLI Demo

24. Dataset for Hands On Lab
/jupyter/data/creditcard/creditcard_train_cat.csv

25. Questions?