Point-in-Time Correct Joins
You'll learn how to join training features with labels using only data that existed at prediction time, preventing subtle data leakage.
When building ML training datasets, you typically join features from your feature store with historical labels. The critical challenge: you must only use feature values that were actually available at the moment you're simulating a prediction, not future data that leaked backward in time.
The Problem
Imagine you're predicting customer churn on January 15th. Your feature store contains a "30-day purchase count" feature that gets updated daily. If you naively join today's feature values with January 15th labels, you might accidentally use the purchase count as it exists today (which includes purchases made after January 15th) — that's data leakage.
How Point-in-Time Joins Work
A point-in-time correct join ensures that for each label timestamp, you retrieve the feature values as they existed at or immediately before that timestamp. It's like asking: "What did we know about this customer on January 15th, not what we know now?"
This requires your feature store to maintain temporal versioning of features. When joining:
- For each label at time
t, fetch features with timestamps<= t - Use the most recent feature value before
t - Never peek into the future
Analogy
Think of it like a history exam where you can only use information from textbooks published before the historical event you're analyzing — not modern textbooks that describe what happened afterward.
Key Takeaway: Point-in-time correct joins prevent temporal data leakage by ensuring training examples use only features that existed at prediction time, creating an honest simulation of production inference conditions.