In other words, I don't have an intuitive explanation on how the logic works, lol.
I put a lot of effort in manually tuning the UMAP hyperparameters and qualitatively assess if it made sense to me. This version is the best I got after ~8hs of fiddling with it.
DanielQuelali
6
Posts
A member registered Jun 12, 2023 · View creator page →
Creator of
Recent community posts
I used `all-MiniLM-L6-v2` embeddings for the top 10k most frequent words, then reduced dimensionality to 2d hyperbolic space by following this doc as reference: https://umap-learn.readthedocs.io/en/latest/embedding_space.html
Then, as shown in the doc, mapped into a Poincaré disk
Thanks for the feedback! I've heard a similar comment from friends. Initially, I considered a square-shaped map, but realized it limited replayability to just 4 words (the vertices). That's why I went with the circle. Adding a terrain background with island names is something I'm considering to improve clarity. Thanks for playing!