Best Practices for ML Engineering by Martin Zinkevich
Matrin Zinkevich from Google recently published and shared the ‘ Rules of Machine Learning’ document which talks detail on Best Practices for Machine Learning Engineering. If you are taken the course in the Machine Learning/Artificial Intelligence, working in ML/AI domain or job seeker, I would strongly recommend to go thru this best practice document.
List of rules on ML Engineering,
· Rule #1: Don’t be afraid to launch a product without machine learning.
· Rule #2: Make metrics design and implementation a priority.
· Rule #3: Choose machine learning over a complex heuristic.
ML Phase I:
Your First Pipeline Rule
· Rule #4: Keep the first model simple and get the infrastructure right.
· Rule #5: Test the infrastructure independently from the machine learning.
· Rule #6: Be careful about dropped data when copying pipelines.
· Rule #7: Turn heuristics into features, or handle them externally.
· Rule #8: Know the freshness requirements of your system.
· Rule #9: Detect problems before exporting models.
· Rule #10: Watch for silent failures.
· Rule #11: Give feature sets owners and documentation.
Your First Objective
· Rule #12: Don’t overthink which objective you choose to directly optimize.
· Rule #13: Choose a simple, observable and attributable metric for your first objective.
· Rule #14: Starting with an interpretable model makes debugging easier.
· Rule #15: Separate Spam Filtering and Quality Ranking in a Policy Layer.
ML Phase II: Feature Engineering
· Rule #16: Plan to launch and iterate.
· Rule #17: Start with directly observed and reported features as opposed to learned features.
· Rule #18: Explore with features of content that generalize across contexts.
· Rule #19: Use very specific features when you can.
· Rule #20: Combine and modify existing features to create new features in humanunderstandable ways.
· Rule #21: The number of feature weights you can learn in a linear model is roughly proportional to the amount of data you have.
· Rule #22: Clean up features you are no longer using.
Human Analysis of the System
· Rule #23: You are not a typical end user.
· Rule #24: Measure the delta between models.
· Rule #25: When choosing models, utilitarian performance trumps predictive power.
· Rule #26: Look for patterns in the measured errors, and create new features.
· Rule #27: Try to quantify observed undesirable behavior.
· Rule #28: Be aware that identical shortterm behavior does not imply identical longterm behavior.
Training Serving Skew
· Rule #29: The best way to make sure that you train like you serve is to save the set of features used at serving time, and then pipe those features to a log to use them at training time.
· Rule #30: Importance weight sampled data, don’t arbitrarily drop it!
· Rule #31: Beware that if you join data from a table at training and serving time, the data in the table may change.
· Rule #32: Reuse code between your training pipeline and your serving pipeline whenever possible.
· Rule #33: If you produce a model based on the data until January 5th, test the model on the data from January 6th and after.
· Rule #34: In binary classification for filtering (such as spam detection or determining interesting emails), make small shortterm sacrifices in performance for very clean data.
· Rule #35: Beware of the inherent skew in ranking problems.
· Rule #36: Avoid feedback loops with positional features.
· Rule #37: Measure Training/Serving Skew.
ML Phase III: Slowed Growth, Optimization Refinement, and Complex Models
· Rule #38: Don’t waste time on new features if unaligned objectives have become the issue.
· Rule #39: Launch decisions will depend upon more than one metric.
· Rule #40: Keep ensembles simple.
· Rule #41: When performance plateaus, look for qualitatively new sources of information to add rather than refining existing signals.
· Rule #42: Don’t expect diversity, personalization, or relevance to be as correlated with popularity as you think they are.
· Rule #43: Your friends tend to be the same across different products. Your interests tend not to be.
Original Paper here,