Detecting Hidden Bias in Datasets Before Models Fail
summary
A language‑learning churn model shows strong offline numbers — 80% precision — yet production recall falls by about 40% recall drop because analytics only includes users with enough events, invisiblizing first‑week churners and later a new “streak” feature changes session patterns. Remedies implemented: include early‑user logs, scheduled retraining, shadow deployments, cohort‑level loss checks, and monitoring with Evidently and Grafana. Hidden biases identified: selection, temporal, pipeline and aggregation; concrete thresholds and retraining cadence are used to detect and fix failures.
AI-written summary · about
chapters (12)
- 0:00Introduction to hidden bias
- 1:25Model failure due to biased data
- 3:54Production failure and early users
- 6:21Investigating data collection flaws
- 9:11Underrepresented new learners
- 10:58Impact of new product feature
- 13:53Temporal bias and retraining
- 17:29Pipeline and aggregation bias
- 20:40Hidden bias taxonomy explained
- 23:51Monitoring and tooling for bias
- 25:39Solutions and continuous improvement
- 27:06Q&A session
related talks
20mBuilding Intelligent Retention Engines: Real-Time Churn Detection
18mReliable Maintenance of Machine Learning Models
33mHidden Biases In Data: Moving Beyond Black Box Models
28mHindsight bias: A machine learning story
25mWhy Good Models Fail After Deployment
27mPreventing churn like a bandit
24mChallenges in building Machine Learning models in production
30mWhen Good Models Go Bad: The damage caused by wayward models and how to prevent it