DS from S: Chapter 11
Machine Learning
Some family things came up and it’s taken me awhile to do this chapter. Should be able to get the next few chapters out sooner rather than later, though.
Thought before starting
Looking forward to this chapter. I’m thinking (once again) that this will be the first chapter where the content is genuinely new (rather than just some new tooling).
Thoughts while reading
Initially pretty conceptual: “what is machine learning?” kind of stuff (it’s even the title of a subsection).
Overfitting & underfitting are pretty familiar concepts for me from physics.
Getting some actual vocabulary I plan to use here; creating a Vocab section.
Definitely learning some new things in this chapter, though most of it is conceptual. Looks like we’ll be building on this soon!
What I learned
Shallow copy a list with
data2 = data[:]!I used to always used some version of
data2 = copy.copy(data)from thecopylibrary
Lots of vocabulary terms
What I liked
Discussion of what can go wrong in splitting training & test data with author’s encounter with one such problem
What I disliked
Other thoughts
Vocabulary terms
Predictive modeling
Data mining
Machine learning
Supervised learning: data is labeled with the correct answers to learn from
Unsupervised learning: data has no annotations; can’t predict the “correct” answers but you might find correlations
Online learning: model continually adjusts itself based on new information
Reinforcement: model gets a signal saying how well its predictions perform
Training data set: to train the model
Validation data set: to choose quality from among candidate models
May not be necessary if you’re not choosing from among candidates
Testing data set: used to judge the quality of the final model
Correctness terms:
True positive: predicted positive & actual positive
False positive (Type 1 error): predicted positive & actual negative
False negative (Type 2 error): predicted negative & actual positive
True negative: predicted negative & actual negative
Precision: how accurate the positive predictions are
Recall: what fraction of positives our model predicted
F1 Score: Harmonic mean of precision and recall
Note: harmonic mean is the the inverse of the mean of the inverses, i.e., average speed for two trips of equal length is the harmonic mean of the speeds.
Bias: error from bad assumptions in the model; you miss the actual connections between features & target outputs
High values can come from underfitting the data, or from other sources
Variance: error from sensitivity to minor variations in the data
High values can come from overfitting the data, or from other sources
Bias-variance tradeoff
High bias & low variance: underfit
Low bias & high variance: overfit
Feature: any input I provide to my model