Time-series Feature Engineering Done Right
summary
Workday builds retention-risk models using a temporal validation pipeline: train on events from 6–3 months back, validate on the last 3 months, and slide or overlap windows to capture seasonality; k-fold gives inflated scores. Feature engineering focuses on time-since-acquisition buckets, org-level attrition rates and derivatives, PTO spikes, manager-change signals; linear models match tree models after these features. Two predictive apps go live—two apps in production for over 100 customers—and survival models prove harder to scale.
AI-written summary · about
chapters (9)
- 0:00Introduction and background
- 2:14Problem formulation and validation
- 11:08Temporal validation methodology
- 17:04Why standard k-fold fails
- 21:01Feature engineering foundation
- 27:22Time-dependent features
- 31:47Data pipeline and production
- 32:31Lessons learned summary
- 33:13Q&A session
related talks
01h 28mFeature Engineering Made Simple
40mCreating and operating ML models from event based data using feature stores
39mFeature Engineering: Unlocking Potential for Predictive Models
21mTime for quality ML from event-based data
01h 52mDeploying a Fraud Detection Model with a Feature Store
43mFeature Engineering for Time Series Forecasting
17mCreating & operating ML models from event-based data using feature stor...
30mFeature Engineering: An Apprentice’s Guide to the “Dark Art” of Machine Learning