safefeat: Leakage-Safe Feature Engineering for Event Logs (Python)
Published:
Tech: Python · feature engineering · machine learning · MLOps
safefeat Tool
safefeat is a Python package for leakage-safe, point-in-time feature engineering from event logs. It builds features for each (entity_id, cutoff_time) pair using only events that occurred at or before the cutoff time, ensuring no future data leakage — a common pitfall in building ML models on temporal data.
The package works around three core components: a Spine (defining when predictions are made), an Events table (historical event log), and a Feature Specification (a declarative description of what features to compute). This design makes it straightforward to build reproducible, production-ready feature pipelines for tasks such as churn prediction, fraud detection, and behavioural modelling.
Key features include:
- Window aggregations — count, sum, mean, and unique counts over configurable time windows (e.g. 7D, 30D)
- Recency features — time since last event, useful for churn, fraud, and behavioural models
- Configurable lag — strict enforcement of point-in-time correctness via
allowed_lag
My Contributions
- Designed and built an open-source Python package for leakage-safe feature engineering from event logs • Published on PyPI • Full MkDocs documentation • Built with applied ML workflows in mind, particularly for user behaviour, fraud, and healthcare event data.