Data Science beasts (failures) and where to find them
summary
Shows concrete failures and fixes: claims 2022 produced 97 zettabytes yet only 20% of analytics yield business outcomes, then catalogs four failure modes—data, infrastructure, implementation, culture—with sharp case studies: a loan model trained only on approved memos that raised delinquency, a COVID model that learned hospital font as a severity signal, UK contact-tracing data truncated by Excel limits (1,048,576 rows), Netflix’s $1M recommender win that wasn’t cost-effective, Microsoft’s Tay derailed in <24 hours. Includes data contracts, premortems, prototyping on real data, and monitoring/model-drift steps.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:09Data scale vs business insights
- 3:41Why DS projects fail
- 5:06Data quality beast
- 9:44Infrastructure beast
- 13:21Scaling challenges
- 17:19Cultural misconceptions
- 22:18Mitigation strategies
- 27:34Premortems and risks
- 33:12Summary & Q&A
related talks
25mWhy Good Models Fail After Deployment
56mData, AI, Hype and Fear - Recipe for Disaster?
19mData science demystified
29mData Science Best Practices : From Proof of Concepts to Production
44mKeep it Clean: Why Bad Data Ruins Projects and How to Fix it
26mThe Hidden Foundation: How Effective Data Management Determines AI System Success
25mBuilding A Successful Data Science Team
45m"Breaking black-box machine learning"