I have the same suspicion, but you're better qualified to answer that question!

This is great! Very intuitive. I suppose pooling will have this effect with any sort of aggregation along an axis? Pretty useful.

Not silly! I don’t think there’s anything wrong with eyeballing it (which is what I did here).

It’s difficult to “average” a set of points because of the same order-invariance issue.

I’ve found success with something called a Wasserstein Barycenter.

That's what I've been playing with! But would prefer some of your feature engineering out of the box!

What I’d really love is if unravelsports could be used to make node-level predictions instead of just frame-level.

One of the first things I want to try with this is plugging in your GNN model.

Big thanks to @probberechts.bsky.social – this uses his lovely 𝗱𝟯-𝘀𝗼𝗰𝗰𝗲𝗿 package under the hood.

And a live demo on Streamlit Community Cloud:

You can find the GitHub repo here.

Really enjoyed the idea of using the sideline to apply equal-and-opposite pressure relative to the player velocity.

Super clever!