Debugging machine-learning My presentation from PyDataWarsaw 2017

1. Debugging Machine Learning.
Mostly for profit but with a bit of fun too!
Michał Łopuszyński
PyData Warsaw, 19.10.2017

2. About me
In my previous professional life, I was a theoretical physicist.
I got a PhD in solid state physics
•
For the last 5 years I work as a Data Scientist in ICM, University of Warsaw•

3. Agenda
4 failure modes of ML systems•
9 simple hints what to do•

4. Hey, my
ML system
does not
work at all...

5. Hint #1
Check your code
AKA it is engineering, stupid!

6. Write tests
Do not strive for 100% coverage, partial coverage is infinitely better
then none!
•
In Python, doctests are your friends•
"Hidden" benefits of tests•
Better code structure•
Executable documentation•
Test the fragile parts first•

7. Test your code with Monte Carlo / synthetic data
Try to generate a trivial case for your
ML system
•
This tests the whole pipeline (transforming/training) and allows for exploration of
your system performance under unusual conditions
•
If you have generative model,
prepare Monte Carlo data from
assumed distributions
•
Perturb the original data, by
generating the output with known
and learnable prescription
•

8. Code style
Single responsibility principle•
Do not repeat yourself (DRY!)•
Have and apply coding standard (PEP8)•
Consider using a linter•
Short functions and methods•
https://xkcd.com/844/
pycodestyle checker (formerly pep8)•
yapf formater•
pylint, pyflakes, pychecker

9. Naming
Minimal requirement - be consistent!•
Interesting software development books, offering chapters on naming•
Freely available chapter on names

10. Hint #2
Check your data

11. Data quality audits are difficult
Happy families are all alike; every unhappy family
is unhappy in its own way.
Leo Tolstoy
Like families, tidy datasets are all alike but every
messy dataset is messy in its own way.
Hadley Wickham
H. Wickham, Tidy Data, JSS 59 (2014), doi: 10.18637/jss.v059.i10
Images credit - Wikipedia

12. Data quality
Beware, your data providers usually
overestimate the data quality
•
Think of outliers, missing values (and how the are represented)•
Understand your data•
Do exploratory data analysis•
Visualize, visualize, visualize•
Talk to the domain expert•
Is your data correct, complete, coherent, stationary (seasonality!), deduplicated,
up-to-date, representative (unbiased)
•

13. OK, my ML system
works, but I think it
should perform
better...

14. Hint #3
Examine your features

15. Features
Features make a difference!•
Be creative with your features•
Try meaningful transformations,
combinations (products/ratios), decorrelation...
•
Think of good representations for non-tabular data
(text, time-series)
•
Make conscious decision about missing values•
Understand what features are important for your model•
Use ML models offering feature ranking•
Use feature selection methods•
ID F1 F2 F3

16. Hint #4
Examine your data points

17. Data points
Find difficult data points! (DDP)•
DDP = notoriously misclassified (or high error) cases
in your cross-validation loop for large variety of models
•
ID
P1
P2
P3
P4
P5
Examine DDPs, understand them!•
In the easiest case, remove DDPs from the dataset
(think outliers, mislabeled examples)
•

18. Influence functions

19. Influence functions
Best Paper Award
ICML 2017

20. Data points
Get more data!• ID
P1
P2
P3
P4
P5
Trick 1. Extend your set with artificial data
E.g., data augmentation in image processing,
SMOTE algorithm for imbalanced datasets
•
Good performance booster, rarely applicable•
Trick 2. Generate automatically noisy labeled
data set by heuristics, e.g. distant supervision in NLP
(requires unlabeled data!)
•
Trick 3. Semisupervised learning methods
self-training and co-training (requires unlabeled data!)
•

21. Hint #5
Examine your model

22. Why my model predicts what it predicts? (philosophical slide)
How do you answer why questions?•
Inspiring homework: watch Richard Feynman, Fun to imagine on magnets (youtube)•

23. Model introspection
You can answer thy why question, only for very simple models
(e.g., linear model, basic decision trees)
•
Sometimes, it is instructive to run such a simple model on your dataset, even
though it does not provide top-level performance
•
You can boost your simple model by feeding it with more advanced
(non-linearly transformed) features
•

24. Complex model introspection
LIME algorithm = Local Interpretable Model-agnostic Explanations
ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (2016)

25. Complex model introspection – practical issues
Lime authors provided open source Python implementation as lime package•
http://eli5.readthedocs.io/en/latest/tutorials•
Another option eli5 package, aimed more generally at model introspection
and explanation (includes lime implementation)
•

26. Visualizing models
Display model in
a data space
Look at collection of
models at once
Explore the process of
model fitting
The ASA Data Science Journal 8, 203 (2015), doi: 10.1002/sam.11271

27. So my ML system
works on test data,
but you tell me it
fails in production?

28. Hint #6
Watch out for overfitting

29. Overfitting
If you torture the data long enough,
it will confess.
Roald Coase

30. Hint #7
Watch out for data leakage

31. Data leakage
Some time ago, I used to thing data leaks are trivial to avoid•
They are not! (Look at number of Kaggle competitions flawed by Data Leakage)•
You may introduce them yourself
E.g. meaningful identifiers, past & future separation in time series
•
You may receive them in the data from your provider•
Good paper•

32. Hint #8
Watch out for covariate shift

33. What is covariate shift?
Training data
y
X

34. What is covariate shift?
Model
Training data
X
y

35. What is covariate shift?
Model
Noiseless reality
Training data
y
X

36. What is covariate shift?
Model
Noiseless reality
Training data
Production data
(Test data)
y
X

37. Covariate shift
Unlike overfitting and data leakage, it is easier to detect•
Method: Try to build classifier differentiating train from production (test).
If you succeed, you very likely have a problem
•
Basic remedy – reweighting data points. Give production-like data higher impact
on your model
•

38. The quality of my
super ML system
deteriorates with
time, really?
Really really?

39. ML system in production
2009: Hooray we can predict flu
epidemics from Google query data!
2014: Hmm... Can we?

40. ML system in production
NIPS 2015 paper

41. Hint #9
Remember monitoring & maintenance
AKA it is engineering again, stupid!
5

42. Thank you!
Questions?
@lopusz