Keynote: Easier Machine Learning Thoughts From Scikit-Learn
summary
Scikit-learn focuses on concrete wins: algorithmic changes that speed real workloads, not vague scaling promises. Examples include randomized PCA (≈10x speedup on many datasets), using the SAGA optimizer for faster, less finicky logistic regression, and a HistGradientBoosting with binned features that yields orders-of-magnitude training speedups. Parallelism relies on joblib delegating to schedulers like Dask to avoid oversubscription. For dirty tables, empirical results show naive imputation can hurt, so use missing indicators or models with native missing support and DirtyCat substring encoding for categorical noise.
AI-written summary · about
chapters (28)
- 0:00Introduction
- 0:20Industry Standards
- 1:03Typical data set
- 1:30Constant size
- 1:51Impact on society
- 2:54Challenges in data science
- 3:48What weve done
- 4:30Why Python
- 5:23Why Numpy
- 7:56Community
- 9:36Better Algorithms
- 11:12Logistic Regression
- 12:43Gradient Boosted Trees
- 14:07Parallel Computing
- 14:55Machine Learning
- 16:01Scheduling
- 16:39Joblib
- 17:06Dusk
- 17:34Python Virtual Machine
- 18:02Scaling
- 18:25Challenges
- 18:42Data
- 18:52Missing Values
- 19:02Imputation
- 19:12Missingness
- 19:40String representations
- 19:56String similarity
- 20:09substring model
related talks
35mScikit-Learn can do THAT?!
01h 17mMachine Learning with Scikit-Learn
39mPushing Cython to its Limits in Scikit-learn
28mEnhancing Machine Learning Workflows with skore
17mMachine Learning at the Speed of Light with Spark
32mKeynote: AI & the stuff built for AI - are they actually useful for data science?
35mIntelligent Ruby: Getting Started with Machine Learning
39mA full Machine learning pipeline in Scikit-learn vs in scala-Spark: pros and cons