#185: Reimagining limited deck clustering

And don't get me wrong, Seventeen Lands data is amazing.

Win rates are a great way of learning the power of the cards in the format, but there are a lot of caveats.

And one of the biggest one that I was struggling for a long time is it misses the sub-color pair diversity.

When you look at the win rate, you either have a generic win rate of a card, let's say we have Rabble Rouser, and then all the Lorehold decks and all the, uh, Prismari decks bundled together, and then you come up with this one number, which is the Rabble Rouser win rate.

And of course, it doesn't account for the fact that probably Rabble Rouser is better in Lorehold decks because it plays into the main mechanic of that.

Also is a creature and, uh, Prismari decks, as we discussed in the last episode, if you remember, ages ago.

Uh, but Prismari decks want to be lighter on creatures.

So putting one creature that is a mana dork is maybe not the best strategy because it will, of course, eat the removal as soon as it's physically possible.

Uh, but also some cards will be better in some types of a build within a color pair and not so good in the other.

And this...

Well, the color pair thing is addressed by Seventeen Lands because you can look at the specific win rate in Lorehold decks only.

But there can be different types of Lorehold decks.

There can be different types of Prismari decks, and you can't extract that kind of win rate looking just at Seventeen Lands data.

So I sat down three weeks ago in order to try and solve this problem.

And I partially did it before when I did my cube analysis episode and I looked at different cube decks because cube is particularly struggling, uh, with this.

Um, but also cube is very easy to sort it out because cube has 540 different cards.

Every pack has one of them.

There's no rarities, which means there's a vast diversity of decks in there.

Um, and because there is a vast diversity of the decks, it's easier to cluster them because that's what we're gonna be talking about, clustering decks.

Um, but clustering on retail sets was something that always eluded me because, um, very often when you look at the...

When you put the decks com- com- compositions into any pipeline that will cluster the decks by composition, you will end up with one big cloud of Boros decks, and it will not have enough power to distinguish the sub-builds of Boros.

So the difficulty for me was to spend the next couple of weeks trying to figure out the clustering algorithms, and there's many of them, and find one that is particularly good with those kind of data samples when there is a lot of zeros.