Tech Talk at Lookout (24 January 2012)

This is a slide deck (for public release) from a tech talk I gave at Lookout on R and a light introduction to unsupervised learning through k-means clustering.

Transcript

1. External Introduction to R (And Machine Learning) Tech Talk, 24 January 2012
2. External Wat What is R? What makes a good model? Example: Modeling flowers Example: Modeling users
3. External Who’s heard of R? Internal
4. External Who’s heard of R? Internal
5. External Who’s used R? Internal
6. External Who hates R?
7. External History of R in 60 seconds Started life as “S”--the FORTRAN-based “statistical programming language”. (1975) R began as an implementation of S with syntax inspired by Scheme. (1993) Core mainly written in C and Fortran.
8. External Why it sucks Legacy stack Written by non-engineers (like me): spaghetti code, difficult to maintain, just, just awful Slow for data-intensive I/O tasks, concurrency and parallelized tasks is akin to skinning alpacas
9. External Why it’s awesome Great for exploratory analysis of raw data Rich statistical library of packages Hooks into “regular” languages easily RPy, rsruby gem, Rtalk
10. External Download R Windows R OS X Native RStudio ESS (Emacs speaks Statistics) Every flavor of Linux R-Vim
11. External What makes a good model?
12. External Good models are
13. External Good models are Precise
14. External Good models are Accurate Precise
15. External Good models are Accurate Precise Generalized
16. External
17. External Example: Modeling TV
18. External All the characters on Friends are Kitty cats. (Not accurate or precise)
19. External white-collar felines between the age of 25-30. (Precise, but not accurate) All the characters on Friends are
20. External humans. (Accurate, but not precise) All the characters on Friends are
21. External 20-something raging hedonists who surely live in a rent-controlled apartment in lower Manhattan. (Accurate and precise) All the characters on Friends are
22. External You see what’s going on here... A model that’s 100% accurate, precise and generalized doesn’t exist! There will always be tradeoffs between all three Your ideal model will surely depend on your risk tolerance for all three Does it need to be fast enough for an MVP? (eg real-time referrer analysis) Does it need to be absolutely accurate? (eg malware classification) Does it need to be ultra-granular? (eg customer segmentation
23. External Example: Modeling Flowers
24. External Iris dataset (R example) The data looks like this: Sepal.Length Sepal.Width Petal.Length Petal.Width Species 1 5.1 3.5 1.4 0.2 setosa 2 4.9 3.0 1.4 0.2 setosa 3 4.7 3.2 1.3 0.2 versicolor 4 4.6 3.1 1.5 0.2 virginica 5 5.0 3.6 1.4 0.2 versicolor 6 5.4 3.9 1.7 0.4 setosa
25. External Iris dataset (R example) Can we predict which species from the four features given? Yes, with the following framework:
26. External Iris dataset (R example) Sepal.Length Sepal.Width Petal.Length Petal.Width Species 1 5.1 3.5 1.4 0.2 setosa 2 4.9 3.0 1.4 0.2 setosa 3 4.7 3.2 1.3 0.2 versicolor 4 4.6 3.1 1.5 0.2 virginica 5 5.0 3.6 1.4 0.2 versicolor 6 15.4 13.9 21.7 7.4 setosa
27. External Poor person’s k-means clustering
28. External
29. External
30. External
31. External
32. External Where K-means fails
33. External K-means in R (three steps) library(caret) data(iris) kmeans(iris[,-c(“species”)], iris[,”species”], 3) (Optimization left as an exercise)
34. External Questions?