How to encode categorical variables in a video game predictive model

Question

I'd like to make a model to predict the result of a match in a video game (win or loss).

The game is 3 players against 3 players, and each player has a specific character with specific characteristics, abilities, etc.

There are around 15 different characters.

My problem is that I don't know how to model that kind of data because it's non-numeric data, and, for example, Scikit-Learn doesn't handle categorical variables.

I read a lot about how to handle that, but couldn't find an approach which seems to be the good one for my case.

I can't really find a numeric value which could be representative and replace the character name variable. Do you have an idea of how I should do that, or maybe just a way to find out?

Answer 1

If the only possible results are win/lose, then you are in front of a binary classification problem (so SVM, trees, logistic regression and all of the others can apply). I don't know about scikit-learn in particular, but most machine learning packages will create the dummy variables for you (you only have to provide them with a column containing the labels, and the software will do the job)

The idea behind these dummy variables is to separate a categorical variable with $n$ groups into $n$ different binary variables (if the individual in question belongs to the $k$th category, then the $k$th dummy variable will be $1$ and the rest will take the value $0$)

In practice, what is most commonly done is take a "reference group" (normally, the majority group) and $(n-1)$ dummy variables. If all dummy variables are $0$, the individual belongs to the reference group. Otherwise, it belongs to the corresponding alternative category