Keep it Clean: Why Bad Data Ruins Projects and How to Fix it
summary
Claims that bad data costs $3.1 trillion and that 27% of data is flawed are backed by concrete failures: COMPAS produces racially skewed recidivism scores, Amazon’s hiring model is discarded for gender bias, Tay quickly learns abuse from live tweets, Babylon retracts accuracy claims, Google Translate flips gender, and Cloud Vision mislabels the Mona Lisa. Practical fixes include visualize everything, drop unusable rows, impute with mean/median or model-based estimates, use secondary sources like Instacart’s GPS, aggregate or exponential-smooth to suppress noise, and watch for data leakage.
AI-written summary · about
chapters (23)
- 0:00Intro
- 1:12The Compass Model
- 6:50Babylon Health
- 7:24AI Doctor
- 10:14News Removed
- 11:45Google Translate
- 16:21Adverse adversarial attacks
- 19:18The numbers
- 20:23Costing money
- 20:48Why this happens
- 21:33Translation
- 21:57Noise
- 23:14Biases
- 25:07Visualization
- 25:42Data Availability
- 27:20Data Consistency
- 28:17Data Leakage
- 29:37Titanic Dataset
- 32:17How to Fix Missing Data
- 35:32Feature Noise
- 38:19Noise Reduction
- 39:39Anomalies
- 41:44Wine Quality
related talks
49mKeep it Clean: Why Bad Data Ruins Projects and How to Fix it
25mWhy Good Models Fail After Deployment
49mData Cleansing & Beyond: How to Address Data Debt for AI
38mSocial Implications of Bias in Machine Learning
31mThe Hidden Costs of Data Quality - Tackling Common Data Challenges in ML
35mBad data - anecdotes and examples from the real world
34mData Science beasts (failures) and where to find them
30mBias, Explainability, and Accountability: The Data Scientist's Burden