Predicting winrate of hearthstone decks using their archetypes

This paper describes our solution for the AAIA'18 Data Mining Challenge: Predicting Win-rates of Hearthstone Decks. Train and test decks were clustered by DBSCAN algorithm with precomputed distance matrix dependent on the number of common cards. We observed that each cluster can be represented by an archetype deck - one of popular decks used by human players. For each deck we created features describing cards quality and types. Additionally we used differences of these features with respect to archetype decks. Finally we used XGBoost to build a model predicting outcome of a game played between two decks.