When I was a teenager I was very popular among my peers, which is why I spent every second Saturday playing in a retiree’s social golf club. For each event we would travel via air-conditioned coach to a local public golf course and play some sort of themed game. The most exciting event was the three stick event. This game requires players to pick just three golf clubs out of their bag (plus a putter) to play their entire round (you can take up to 14 clubs on a normal round).
So which three clubs do you pick?
This was easy for me as I could only reliably hit two clubs: the five iron and the seven iron. I usually included a driver so as not be mocked by the septuagenarians at the par-5 tees.
I also have an embarrassingly crap collection of golf clubs as many have been sent to an early retirement in water features.
However, many friends who now play golf have a bewildering selection of exotic clubs from hybrid driving irons to 60-degree lob-wedges.
It seems like in there is always some new technology or method to master in data science broadly. This is interesting, but also exhausting and fills me with existential dread that I am becoming irrelevant.
It got me thinking, if you had to play statistics (or data science) for the rest of your career with just three algorithms/models, what would they be?
My three sticks:
Random Forest: Uncontroversial ensemble method that can quickly get a performance ceiling with minimal fuss. It lacks some interpretability, but I’ll keep it in my bag for when I want to squeeze as much AUC as I can in a prediction setting.
Generalised Additive Model (GAM): You could argue this is quite broad, but that’s okay. I’ve had good luck with GAMs as they provide a solid statistical approach that I am comfortable with, along with the flexibility of using smooths for non-linear terms that seems to be a key feature in the problems I have to solve. If I had to only pick one club this would probably be it.
k-medoids clustering: Clustering methods are a great third club to have in your bag. They solve a different problem to the modelling tasks above and provide a good hedge for those unsupervised tasks that come along fairly regularly. While everyone knows k-means, I enjoy the increased interpretation of k-medoids as the cluster centers are always one of the input data points. It also allows for more generalised distance methods (such as Gower’s distance) which I have found useful for mixed data types.
We can include lm() as the default putter you get for free.
Which three clubs would you pick in your bag?