Card2vec is a proof-of-concept tool for finding similar cards using a specific machine learning technique.

Start by typing @ in the search bar, then use autocomplete to select a card and hit enter to search.

Add more cards to your search to find matches for several cards.

Remove cards from the search bar with backspace or by hovering over them and clicking the 'x'.

Return to this help page at any time by clicking the card2vec logo.

The similarity score and bar at the bottom of each card in the search results shows how similar cards are to the cards in the search.

Touch or mouse over the "concept bars" to the right of each card to view any "concepts" that may be related to the search.

Card2vec is based on a famous technique in natural language processing known as "word2vec" [1] [2]. The technique uses large amounts of text to compute numerical representations (vectors) of words such that the vectors for similar words are similar.

Card2vec applies the same technique to MTG decklists to compute vector representations of cards, which can then be mathematically compared for similarity.

See this GitHub readme for more details on card2vec.

You can read more about how word2vec works here.

These are determined and labelled algorithmically, and are based on the 5 cards "closest" to a point in vector space. "Concepts" are generated using k-means clustering, using the vector at the cluster center. They are labeled using the openai api (i.e., chat-completions). So if they don't make sense, blame chatGPT.

Approximatey 3.6 million decks were scraped from 6 MTG sites in April of 2023, specifically for this project.

Card images and metadata were retreived through the Scryfall API.

I used the python library Gensim's word2vec implementation.