Item — HN Reader

Completely agree with this. To really learn linear regression, you need to be comfortable with linear algebra (for OLS) and calculus (gradient descent).

I actually wrote most of the machine learning content at Dataquest (where I work) and I started with k-nearest neighbors because it's way more approachable (https://www.dataquest.io/course/machine-learning-fundamental...). Little math, very visual, easy to program, etc. I used this easier-to-approach algorithm to teach the key other ideas in ML (test/train, cross validation, error metrics, etc).

Happy to chat more about different pedagogical approaches to teach machine learning!

2 comments

> I actually wrote most of the machine learning content at Dataquest (where I work) and I started with k-nearest neighbors because it's way more approachable

That's a great idea. It's actually what I did in an earlier version of the USF course. It worked great. Especially because a decision tree is basically just KNN with a different distance measure (loosely defined) so it can flow well

However it turned out that jumping straight to decision trees worked out well too, so I'm happy with the change.

Makes a lot of sense actually. Decision trees are also incredibly visual and a lot more intuitive to understand.

>Completely agree with this. To really learn linear regression, you need to be comfortable with linear algebra (for OLS) and calculus (gradient descent).

There is no need to know gradient descent to learn OLS. All you need to be able to is take derivatives (and "matrix" derivatives)