DS from S: Chapter 13
Naive Bayes
Thought before starting
Really looking forward to getting some practice with Bayes’ Theorem. I didn’t encounter it in any of my statistics courses, which I think was an oversight in the curriculum that kind of made sense, since so much of the statistics in physics can be treated as a normal distribution kind of thing, and because it’s often easy to design most undergraduate labs in such a way that the cause of an experiment’s outcome is clear in physics.
As an aside, the causal relationship of the measurement I spent my graduate years doing was much less clear. It was a measurement of the gravitational constant, big G, and it’s the least well known of the physical constants. Currently we know the value of G to around 22 parts per million; the mass of an electron is known to about 300 parts per trillion, so 0.0003 parts per million, or about 70,000 times more precise. Some of that is caused by the experimental precision, but part of it comes from the fact that it’s easy to overestimate how good your experiment is. There are experiments with less than 22 parts per million uncertainty, but they disagree with each other by more than that. My experiment, by the way, was not one of them. We should have had a more precise experiment than that, but when we changed something that shouldn’t have mattered, our value was different. My whole thesis was basically a survey of things that weren’t the cause.
OK, on to Bayes!
Thoughts while reading
* Sorry for the asterisks instead of bullets, Substack isn’t working well on my browser at the moment (Firefox). I’ve turned off all of my extensions for this page but that doesn’t seem to be fixing the problem.
* Oh nice, we’re using the fake social network introduced in chapter 1 that we haven’t seen since. Or if we have, I no longer recall it.
What I learned
How to write a naive Bayes filter!
What I liked
Good example to work with, and I enjoyed checking out the dataset
What I disliked
Other thoughts
I’d have liked it if the author explained what choice for
Vocabulary terms
Pseudocount: modified count of a value used in a naive Bayes filter to ensure that something that is 100% indicative of whether a term is in one category or another in the test set doesn’t result in the model assigning the word a 100% predictive ability