This repo contains the source for processing and visualisazing hearthstone matchs data. The final map is based on 200 000 ranked hearthstone games. This coressponds to more than 4 000 000 cards played and recorded by collecto-bot over one year and an half.
A live version of the map is available here.
Using this dataset we constructed 400 000 series of played cards one for each player in a game. These series were fed to a word2vec embedding [Mikolov, Tomas et al] algorithm (with negative sampling [Goldberg, Yoav; Levy, Omer] and a single soft-plus final layer) to learn 64 dimensional vectors for embeding all the 1600 cards. The final 2d projection was derived from these vectors using a t-sne [Van der Maaten, Laurens ; Hinton, Geoffrey ] with default settings. The code is available on github. The same dataset was used to estimate winrates and frequency of each card.