[oneoff786]

Hold up. They did this on beat saber… with a 50000 class output (one per user).

They seem to be using mod songs, and are not accounting for the fact that people are playing different songs.

Your motion is entirely determined by the song. I think they’re just identifying pairs of song choice and user height.

Note that one of their metadata features is “replay > 100”, meaning they probably have examples of the same user doing the same song many many times

[koolala]

Seems like a good way to validate it would be to have everyone play the same set of songs. Then to tell if the predictions work, see if it can find each different song each individual plays. I hope the design wasn't as bad as you say.

[oneoff786]

Reading into it more, the second most predictive feature was a context variable. That is, identifying features of the specific song.

I think this one gets a thumbs down from me.