Machine Learning in Production 101
I just finished some writing for the UN (ITU) about machine learning models in production.
Wonder how to deal with data changing and models doing worse after training? (when deployed)
This is for you.
📌 We're talking about Drift
Our training data is static. Contact with the real world is non-stationary.
This drift can happen in three ways:
- The input data changes
- The labels for the data change
- The inherent relationship changes
⚗️ Input data changes!
One way to monitor these is by checking the distribution of the new data vs the training data.
We can use these tests:
- Continuous: Kolmogorov-Smirnov test
- Categorical: Chi-squared test
Solution:
Retraining the models regularly.
🧩 Target / Label changes
These can be natural changes similar to the input changes.
In that case, you can use the same approach.
But sometimes, our categories change, because we make discoveries or mgmt decisions.
Solution: Updates are best reflected in automated pipelines
🔀 Concept shift
This one sucks.
ML models learn the relationship between input and label (ideally).
When that relationship changes our entire historic data set is obsolete.
Essentially what happened in early 2020.
Solution: New data, but setting up auto alerts is essential
📖 More info?
I wrote an ebook about machine learning validation.
I give it away to my newsletter subscribers.
I have just made the biggest update to the ebook, including production models and machine learning drift.
Subscribe to receive insights from Late to the Party on machine learning, data science, and Python every Friday.
Conclusion
We hope training data represents real-world data in machine learning, but it doesn't always.
- Set up MLOps automation
- Retrain for input data changes
- Care for label changes
- Hope it's not concept drift, where the relationship of data changes