Prediction problem across a wide space

I have assembled a database of Clash Royale games in an attempt to understand the outcomes of various match-ups. The game is composed of an 8 card deck drawn from 102 cards. As you can see from the Cnr, it’s a very wide space with many possibilities. Similarly, this deck will be opposed by a wide and diverse space.

I have been using essentially a 206 element wide table. One of the variables relates to the relative strength of the players; one hundred and two variables are indicators for the "home deck"; one hundred and two variables are indicators for the "away deck"; and the final element is an outcome variable that is the binary response variable of interest.

I have utilized an xgboost model to predict the outcome of the opposing decks, but I understand that it is not that useful for extrapolation where card combinations have not actually been played in the data. It seems to me that a logistic approach would be better, but the challenge i am faced with is that the interactions are highly critical, but also explode into too many features to manage.

As additional information, I have used segmentation to "group" the decks into commonly utilized "metas" with good success. Per my note above, however, this does not help in the identification of "new metas" that might be worth exploring.

I am interested in suggested approaches to solving this problem. And if you care to see some of my Clash Royale data, I host a website for my clan at The6ixclan.ca