Tools and Tricks from a Pragmatic Data Scientist
summary
Claims concrete wins: cosine→Euclidean via normalization lets KD/ball‑trees return exact kNN results (no approximation) and empirical timings on MNIST and up to 40k items show dramatic speedups over brute force; missing features get handled by a law‑of‑total‑probability reduction assuming conditional independence, improving classification on degraded digit tests versus simple imputations; heavy‑tailed recursive binning produces interpretable breaks and raises Boston housing R2/RMSE by ~6%; all code uses Python and scikit‑learn and sits on github.
AI-written summary · about
chapters (6)
- 0:00Introduction
- 1:05KNN math for speed
- 12:08Bad data classification
- 22:15Scottish discretization
- 32:10Pragmatic principles
- 34:44Q&A
related talks
53mRobust Algorithms for Machine Learning
42mDistribution-Specific Analysis of Nearest Neighbor Search and Classification
21mKeynote: Easier Machine Learning Thoughts From Scikit-Learn
35mIntelligent Ruby: Getting Started with Machine Learning
34mDimensionally reducing data squeezing out the good stuff
48mCommon Data Science Mistakes
44mWinning with Simple, even Linear, Models
37mOptimal on Paper, Broken in Reality