Skrub DataOps, Probabl
summary
Claims that messy data science code yields 88% of projects failing to reach production and that the fix is better primitives, not new models. Demonstrates Skrub DataOps with a computational graph that records pandas-like transforms, a table vectorizer plus pipeline that preserves fit/transform separation, cross-validated grid search finding an R2 ≈ 0.84 baseline, and a Score system for experiment tracking. Shows replayable multi-table assembly, automatic tuning, and a working demo run locally.
AI-written summary · about
chapters (8)
- 0:00The problem with messy code
- 0:50Coding a predictive model
- 5:14Challenges in data preparation
- 9:07Introducing skrub
- 12:47Skrub DataOps and automation
- 16:46Benefits of skrub primitives
- 17:22Experiment tracking with skore
- 18:13The future of data pipelines
related talks
19mSkrub: Less data wrangling, more machine learning
26mSkrub: machine learning for dataframes
11mBig Data Moonshots and Ground Control
35mApplied Data Science
41mBetter Code for Data Science
01h 14mUp your Bus Number - A Reproducible Data Science Workflow
28mThe Coding Portion of Data Science
36mRun a benchmark they said. It will be fun they said.