The document discusses adding machine learning capabilities to a web application. It notes that adding machine learning is not as simple as it may seem at first and outlines three key steps: 1) Get and explore data, 2) Frame the problem by defining what success looks like, and 3) Measure performance honestly. It provides examples of using Bayesian filtering to suggest tags as users type tasks and discusses piloting models with offline data before full implementation.
2. It’s easy, right?
1. Get some data
2. Find magic stats & algorithms
3. Predict the future!
3. It’s easy, right?
that helps solve the problem
1. Get some data
frame the problem explore the data
2. Find magic stats & algorithms
what’s success? pilot
3. Predict the future!
does this help users? behaviour changes?
UI impact?
12. A contrived example
#home 40% #work 60%
“windows” “fix” “windows” “fix”
50% 50% 16.6% 83.4%
p(#home | “fix”) = 40% x 50%
(40% x 50%) + (83.4% x 60%)
= about 28%
13. A contrived example
p (#h ome | “f i x ” )
#home 28% #work 72%
“windows” “fix” “windows” “fix”
50% 50% 16.6% 83.4%
p(#home | “fix”, “windows”) = 28% x 50%
(28% x 50%) + (16.6% x 72%)
= about 55%
15. p(C | e) = P(C) x P(e | C)
P(e)
“the estimation of P(e | C) can be viewed as the
central issue in designing learning systems. ”
— Weiss & Kulikowski
“Computer Systems that Learn”